Start Here

Autonomy note: docs/RESEARCH_AUTONOMY_CHARTER_DRAFT.md is the byte-identical approved charter snapshot, and docs/RESEARCH_AUTONOMY_CHARTER_DECISION.md is its activation record. Tier A routine work and Tier B bounded development experiments may proceed autonomously after that decision is remotely green; Tier C irreversible evidence, real-data, hardware, destructive, release, and claim actions still require separate exact permission.

Scientific strategy, updated 2026-08-30: read docs/SCIENTIFIC_DISCOVERY_AND_INVENTION_CONSTITUTION.md, then docs/SCIENTIFIC_CONVERGENCE_AND_INVENTION_PLAN.md, then registries/scientific_knowledge_ledger.v0.json. Dreyer and Ofner are parked predecessors. The former prospective question was OFNER-C6R-1: public metadata selected 15 participants / 150 original motor-imagery GDF files / 13,748,417,608 bytes with reported 61 EEG, three EOG, and 32 hand/arm movement channels. Its generated selector and header parser passed their synthetic qualifications. The sole real range-header invocation then consumed at OFNER-H0-TRANSPORT: manifest identity selection completed, but the first 256-byte GDF range yielded zero accepted GDF body bytes. No real header, channel roster, geometry, signal, event, target, model, prediction, or score was observed. Runtime was 1.1932796670589596 seconds, peak RSS was 48,021,504 bytes, and no payload was retained. This is not a biological result. Never retry, rerun, repair, resume, substitute, reinterpret, or reopen its ignored marker. No Tier C packet is active. The exact Ofner proof closeout passed Base Python 99174411928, Optional Neuro Readers 99174412006, and CI 33280371097 on GitHub main. Generated-only NPA1-G then passed at exact commit 2e164ff, Base Python 99184746988, Optional Neuro Readers 99184747065, and CI 33284320443: two matching seven-profile replays and 37 refusal families, 0.003375791013240814 seconds, 22,577,152-byte peak RSS, zero network, and zero retained payload. The NPA1 proof closeout then passed Base Python 99188620896, Optional Neuro Readers 99188621003, and CI 33285776358 at exact commit 2ec3d4b on GitHub main. Read registries/current_research_frontier.v23.json and docs/SCIENTIFIC_RESULT_HARD_DECISION_TREE.md; its predecessor registries/current_research_frontier.v21.json, earlier predecessors registries/current_research_frontier.v20.json, registries/current_research_frontier.v19.json and registries/current_research_frontier.v18.json, earlier predecessor registries/current_research_frontier.v17.json, earlier predecessor registries/current_research_frontier.v16.json, earlier predecessor registries/current_research_frontier.v15.json, earlier predecessor registries/current_research_frontier.v14.json, earlier predecessor registries/current_research_frontier.v13.json, predecessor registries/current_research_frontier.v12.json, predecessor registries/current_research_frontier.v11.json, predecessor registries/current_research_frontier.v10.json, predecessor registries/current_research_frontier.v9.json, predecessor registries/current_research_frontier.v8.json, predecessor registries/current_research_frontier.v7.json, and historical registries/current_research_frontier.v6.json remain immutable. No fresh source is promoted. The pushed FMSR1-v0 draft was rejected before acceptance or protected work. Corrected FMSR1-v1 is exact-green at e09f6cc, Base 99197577034, Optional 99197577007, and CI 33289147031. It freezes five official source indexes, four exact queries, at least ten complete participants, EEG/EOG/EMG for every relevant effector, named geometry, and joint nuisance-only and joint nuisance plus deranged-central-EEG controls. Its 20 GiB ceiling is partitioned exactly as 12+2+2+1+3 GiB. Proof closeout and all-false FMSR1-DISCOVERY-M0 request are exact-green at 5c466efd, Base 99204462828, Optional 99204462714, and CI 33291744779. The packet-bound FMSR1-DISCOVERY-M0-D0 decision is exact-green at b66a3a1, Base 99316322973, Optional 99316322922, and CI 33333549395 on GitHub main. Generated-only FMSR1-DISCOVERY-M0-G1 is exact-green at e92e9c2, Base 99327605264, Optional 99327605346, and CI 33337741165. It passes two deterministic 17-root mock-HTTP replays and 25 adversarial refusals in 0.0161428339779377 seconds at 52,510,720-byte peak RSS, with zero real network, payload, model, score, or claim operations. Independent review found no packet-bound official-index revisions and no external authentication for a self-reported CI record, so the implementation exposes no real transport and the live route is parked before contact. Its proof-only closeout binds six exact artifacts / 162,295 bytes. Exact closeout fac60bf passed Base 99330615390, Optional 99330615564, and CI 33338847448 on GitHub main; it cannot activate live work. Read the all-false A1 admission correction next. It keeps M0 non-armable, leaves all five index profiles and the exact live-CI profile null, and separates future witness, proof, and discovery stages. Exact correction state 8fe98df passed Base 99335260650, Optional 99335260756, and CI 33340527773 on GitHub main; branch CI 33340523548 also passed both jobs. Generated-only FMSR1-R1-G-v0 registration is now exact-green at d53f3e8, Base 99339083749, Optional 99339083636, and CI 33341954248. Read its implementation, durability amendment, evidence-map clarification, and frontier v21 next. Implementation c3f536d passed Base 99347978123, Optional 99347978190, and CI 33345237136, but post-green review rejected activation because directory ancestry durability and pre-arm versus armed consumption were not exact. Corrected I1 d9229c1 and activation 3de14c3 passed their required jobs before the sole generated qualification. That run is consumed and passed with two matching five-profile replays, 82/82 refusals, 0.022029167041182518 seconds runtime, 21,397,504-byte peak RSS, and zero network or protected operations. Result 9f5ff18 and proof closeout dbfaf68 are exact-green; CI 33352882661 passed Base 99369471021 and Optional 99369470875. Never rerun the qualification. The all-false FMSR1-R1-W-v0 packet is exact-green at d4ae388, Base 99381828165, Optional 99381828024, and CI 33357313608 on GitHub main. It binds five opaque snapshot profiles and 17 exact roots in a canonical page/root/profile/global hash tree but grants no authority. Implementation-only decision FMSR1-R1-W-I0-D0 is exact-green at e158e8c, Base 99385124402, Optional 99385124488, and CI 33358495852 on GitHub main. Generated implementation FMSR1-R1-W-I0 then passed its sole fixture qualification: two identical five-profile, 17-root, 34-page replays plus 22 fail-closed mutations in 0.035018666880205274 seconds at 21,364,736-byte peak RSS, with zero network, candidate-semantic, neural, target, model, prediction, score, or claim operations. It is pending this exact implementation commit and remote proof and must not be rerun. One live witness still requires fresh second words, an exact-code-bound second decision, and the three-call commit-scoped GitHub proof. D1 keeps its own later packet/decision barrier. All public network, payload, signal, target, model, score, release, and claim authority remains false. COMM-P0-G FS3 remains paused.

Dreyer H-L1R1 recovery update, 2026-08-29: exact generated coordinator 0ef634e4852d9f8a18d4b8a1f50e6a1331bd020a passed Base Python 99103540731, Optional Neuro Readers 99103540620, and CI 33253657120 on GitHub main. The sole Q0 attempt then passed all 65 frozen generated cases in 0.1789909170474857 seconds at 37,699,584-byte peak RSS, with 1,623,762 generated input-plus-output bytes. It is consumed and cannot be rerun. Exact result closeout cb7e7f831a1056dfcf878ca23ed53aa6fd4738dc passed Base Python 99105690584, Optional Neuro Readers 99105690672, and CI 33254474440 on GitHub main. Exact proof 1ae340354352d544d6d99fe5af6f354ab668bf9c then passed Base Python 99107500593, Optional Neuro Readers 99107500410, and CI 33255170805. The all-false H-L2 request a97fc191106e5fe42859d871d78e59930bef79ac then passed Base Python 99109482999, Optional Neuro Readers 99109482995, and CI 33255920346 on GitHub main. Exact proof 036a9ec7e78b460d464c3349151ddb0e35914d87 passed Base Python 99112795545, Optional Neuro Readers 99112795525, and CI 33257159816. The maintainer's fresh continue is now recorded in packet-bound decision DREYER-C5R-1-HL2-A0-D0. Exact decision 53f2c48831d1db7875bf09a35f107e35f97c6bf4 passed Base Python 99114895023, Optional Neuro Readers 99114895128, and CI 33257975186 on GitHub main. The additive adapter now passes a 32-case / 33-attempt generated qualification with two matching H1 replays and 31 refusal observations in 0.599492667010054 seconds at 39,141,376-byte peak RSS. It used zero network or real/private operations and retained zero generated bytes. Exact implementation a9cd0be7c22996154c28bb568e05c623606e7424 passed Base Python 99121591361, Optional Neuro Readers 99121591482, and CI 33260534900 on GitHub main. No-authority activation DREYER-C5R-1-HL2-ACT0 bound six exact implementation artifacts / 79,741 bytes. Activation e3adab26e867f6de7e56f406f59ede51718b577c passed Base Python 99123725366, Optional Neuro Readers 99123725302, and CI 33261346763 on GitHub main. The sole invocation then refused with HL2-TRANSPORT before body streaming: one GET/response open, zero network body bytes, zero EDF header reads, 0.6479716249741614 seconds, 36,454,400-byte peak RSS, and no payload retained. Never rerun or reopen its ignored evidence.

COMM-P0-G FS3 registration frontier, 2026-08-28: Amendment 3 replaces only the redundant second full fictional model replay with a separate zero-model verifier/scorer. One exact full producer still executes the complete frozen 42-participant schedule; the verifier independently checks its streamed transcript, freeze HMAC, inventory, target-delivery nonce, penalties, and aggregate score. Exact Amendment 3 b15d3cf passed Base 98860211722, Optional Neuro Readers 98860211949, and CI 33174711145, then reached GitHub main. The separately prospective COMM-P0-G-FS3-v0 contract keeps one thread, 180 seconds, 512 MiB RSS, 513 MiB disk, zero network, and zero retained payload. Registration and implementation are exact-green on GitHub main; execution remains false until its remaining green barriers. This is generated engineering only; no scientific claim or Tier C authority changed.

COMM-P0-G FS3 implementation update, 2026-08-28: the additive descriptor- only verifier and reduced coordinator passed two isolated three-person-per- cohort producer/verifier pairs. Each producer emitted 13,056 streamed rows / 204 sets; both canonical producer surfaces and verifier aggregate hashes matched. The run exercised 140 refusals in 72.67009999998845 seconds at 239,075,328-byte peak RSS with 770 monitor samples, zero network, and zero retained payload. Exact implementation a3b561b passed Base 98875866699, Optional Neuro Readers 98875866417, and CI 33179247000, then reached GitHub main. Exact proof-only closeout ccabfaf passed Base 98881035296, Optional Neuro Readers 98881035573, and CI 33180751900, then reached main; it binds six exact artifacts / 43,587 bytes and repeats no execution. No full FS3 or official run has run.

COMM-P0-G FS3 verifier correction, 2026-08-28: independent review found that the frozen score worker makes two sequential stream traversals, while the contract called this one pass, and that exact full cardinality, identity, no-follow, socket-guard, logical cohort-envelope, and runtime semantics needed to be explicit. Prospective COMM-P0-G-FS3-A1 fixes those fields with zero numerical operations. Exact Amendment 1 28b440e passed Base 98889419386, Optional Neuro Readers 98889419758, and CI 33183200744, then reached GitHub main. Exact wrapper 91743f5 passed CI 33188950787, and exact proof 66ab07d passed CI 33190237036; both reached main. The wrapper enforces the corrected descriptor, identity, socket, observed per-set cardinality, resource, and staged-publication semantics. It remains fail- closed only pending a separate activation commit binding proof digest 128b6699; no full FS3 run is authorized or completed.

COMM-P0-G coordinator correction, 2026-08-28: post-proof audit found that the V1 full-scale rehearsal used one combined target envelope and one score, while the frozen contract requires separate discovery and independent- replication delivery and scoring. The generated-only successor now enforces two target deliveries and two scores per replay and implements the complete activation-locked coordinator through two distinct sanitized replay children. Its score freeze is durable and verified before target delivery; every model, shortcut, process, memory, input, output, and disk ledger is checked; and publication is atomic and non-replacing. One reduced two-child measurement passed in 44.6720 seconds at 236,437,504-byte peak RSS with zero network, real-data, device, or retained-payload operations. The compile-time lock is still false. Exact head 669c858 passed Base job 98815336723, Optional Neuro Readers job 98815336577, and CI 33161053075, then reached GitHub main. Read the proof-only closeout. The later FS2 run showed that two full 21-person-per-cohort replays do not fit the exact 180-second cap on this machine: it completed one child and parked at 180.05074683297426 seconds. No official run or claim occurred.

COMM-P0-G full-scale rehearsal result, 2026-08-28: the sole generated-only COMM-P0-G-FS2-R0 attempt is consumed at FS2_PARK. One of two full fictional replay children completed; total runtime reached 180.05074683297426 seconds against the frozen 180-second cap. Peak RSS was 305,119,232 bytes, the exact 537,919,488-byte reservation left 37,325,271,040 bytes free, 1,510 monitor samples were collected, and cleanup retained zero payload bytes. Network, official, real/private, human/device, release, and scientific operations were zero. Exact closeout a654541 passed both required jobs in CI 33171818869 and reached GitHub main. No retry or repair is allowed; official qualification remains inactive.

COMM-LIVE-G0 closeout, 2026-08-27: exact implementation bc9bb109c9c82b56afe06d983d82b5b8ece669cf passed Base 98694674092, Optional Neuro Readers 98694674289, and CI 33123115516, and is present on GitHub main. Proof d7c62b8 passed CI 33123739110. The sole official fictional qualification then passed and was consumed at COMM-LIVE-G0-R1: two byte-equivalent replays, four positive commits, all 33 refusals, 6.421 seconds, and 30,949,376-byte peak RSS. No real EEG, target, model, provider, stream, device, or scientific operation occurred; no rerun is allowed.

Prospective communication cohort, 2026-08-27: after the registered public replication routes failed source qualification, COMM-P0-SYNC-v0 froze a zero-operation two-cohort study with 42 planned completers and a 44-person enrollment cap, synchronized 64 EEG + 4 EOG + 4 oral-EMG channels, microphone and triggers, prompted and free-choice four-command endpoints, participant-held-out inference, causal shadow/live stages, and a 10,463,692,800-byte worst-case raw budget. Read docs/COMMUNICATION_EEG_PROSPECTIVE_SYNCHRONIZED_COHORT_PREREGISTRATION.md. It authorizes no recruitment, recording, hardware, data, model, target, score, release, or claim operation. Exact registration df3266e passed both required jobs in CI 33134791405 and reached GitHub main. Read its proof-only closeout and Amendment 1: review found a frozen 60% human versus 70% machine stable-commit coverage mismatch, so the stricter 70% is authoritative. Generated design remains paused until both additions are remotely green. Exact proof/amendment 478d31e passed both required jobs in CI 33135742217 and reached GitHub main. The next gate is the all-false generated-only COMM-P0-G-v0 registration. Exact registration 002128b passed both required jobs in CI 33136788477 and reached GitHub main. Its proof-only closeout binds three immutable artifacts / 47,562 bytes. Exact proof eafe1df passed both required jobs in CI 33137998642 and reached GitHub main. The generated-only protocol core now represents the complete fictional cohort/trial plan, synchronized 73-role sensor bundle, target precommitments, participant-first statistic, live accounting, and all 70 refusal wrappers. The compact model/two-replay coordinator is now implemented; official qualification remains separately gated. Exact core 9378421 passed both required jobs in CI 33139382019 and reached GitHub main. The compact 42-fold numerical schedule and aggregate-only scorer are now implemented additively. Exact coordinator head 669c858 is remotely green and on GitHub main; it adds distinct sanitized replay children, the complete malformed-refusal matrix, durable freeze verification, and bounded atomic publication. Its full-scale FS2 runtime proof is now consumed and parked; activation remains closed. Exact numerical/scorer b7e4d83 passed both required jobs in CI 33140872221 and reached GitHub main. Exact endpoint-separated live scorer c46064d passed both required jobs in CI 33141715948 and also reached main; free-choice live performance is now primary and prompted performance cannot rescue it. The additive runner scaffold has completed a reduced two-replay development pass at 187,645,952-byte peak RSS with zero retained payload. Exact green 7d625a9 on GitHub main adds the standalone bounded streaming scorer. The V1 hardening successor now streams canonical fold predictions with a one-row raw high-water mark, executes all seven numerical shortcut routes, wires the durable future consumed marker before official work, and completed one nonofficial 42-person fictional rehearsal: 91,392 prediction rows / 1,428 sets in 101.5826 seconds at 283,394,048-byte peak RSS and 34,259,372 private temporary bytes, with zero network, real-data, device, or retained-payload operations. Exact V1 55e627d passed Base job 98789718192, Optional Neuro Readers job 98789718367, and CI 33153174019, then reached GitHub main. Read its proof-only closeout, which binds the exact implementation and rehearsal evidence without repeating any execution. The official two-replay qualification remains separately activation-locked and scientifically inert. The later two-child FS2 attempt is also consumed: one child completed before the shared 180-second deadline, so runtime architecture, not more model capacity, is the next reversible engineering question.

Latest Irreversible Result: Ofner Range Header Consumed At Transport H0

Historical exact control-plane label retained for proof compatibility: Current Gate: DREYER-C5R-1-HL One-File Sensor Preflight.

Start with registries/current_research_frontier.v18.json, the machine-readable snapshot of the consumed lanes, green NPA1 engineering proof, claim boundary, rejected v0 draft, exact-green v1 registration, discovery decision and generated implementation, green proof-only closeout, all-false A1 correction, generated R1-G registration, rejected first implementation, corrected local durability successor, and next transition. Versions 17, 16, 15, 14, 13, 12, 11, 10, 9, and 7 remain immutable predecessors.

The prior BNCI experiment found real held-out protocol information, but it did not establish an EEG-specific effect: posterior EEG slightly beat the selected candidate, candidate probabilities were poorly calibrated, and the small EEG increment over recorded EOG was inconsistent across participants. Read the aggregate-only result first:

docs/BNCI_2014_001_ARTIFACT_POSTMORTEM_RESULT.md;
registries/bnci_2014_001_artifact_postmortem_result.v0.json; and
tests/test_bnci_2014_001_artifact_postmortem_result.py.

The fresh independent replication is now frozen around those failures. Dreyer Dataset A contributes 60 new participants, 27 reported EEG channels, three reported EOG channels, two reported wrist-EMG channels, 512 Hz acquisition, and 120 selected R1/R2 source EDFs totaling 1,779,763,388 declared bytes. The 60-fold protocol asks whether central EEG residualized against EOG, EMG, posterior EEG, and timing adds information in a completely unseen person.

Read, in order:

docs/DREYER_2023_DATASET_A_REPLICATION_PRIMARY_SOURCE_RESEARCH.md;
docs/DREYER_C5R_1_PREREGISTRATION.md;
registries/dreyer_c5r_1_contract.v0.json;
docs/DREYER_C5R_1_GENERATED_QUALIFICATION_RESULT.md;
docs/DREYER_C5R_1_STAGE_H_GENERATED_QUALIFICATION_RESULT.md;
docs/DREYER_C5R_1_STAGE_H_LIVE_PREFLIGHT_AUTHORIZATION_PACKET.md;
docs/DREYER_C5R_1_STAGE_H_LIVE_PREFLIGHT_REQUEST_PROOF_CLOSEOUT.md; and
docs/DREYER_C5R_1_STAGE_H_LIVE_IMPLEMENTATION_SAFETY_REVIEW.md;
docs/DREYER_C5R_1_STAGE_H_LIVE_RECOVERY_QUALIFICATION_IMPLEMENTATION.md; and
docs/DREYER_C5R_1_STAGE_H_LIVE_RECOVERY_QUALIFICATION_CLOSEOUT.md.

The full generated firewall/model/scorer passed once, and the streaming fixed-header preflight passed two valid replays plus 18 adversarial refusals. The first H-L1 wrapper qualification was rejected after finding a cleanup defect. Its additive H-L1R1 successor has now passed the complete 65-case matrix and is consumed. All are generated engineering with no scientific value. The sole later H-L2 endpoint transaction opened one response and then refused before any response body or EDF header was read. No annotation, signal, target, model outcome, prediction, or score was accessed.

The static live-implementation safety review closes the known path-race, atomic-promotion, transport, resource, exception-privacy, and EDF header-to-payload geometry requirements before H-L1. It performs no live operation and changes no authority.

All-false request 5191e79e2ccfa50b5042a24357c0c22d68d8f088 passed CI 32934958878. Its proof-only closeout 821fad17e06914375c50a7d0dd7017458b2df838 passed Base job 98077895278, Optional Neuro Readers job 98077895460, and CI 32936247679. At that historical point DREYER-C5R-1-HL was the sole active Tier C lane. Its H-L1R1 result and proof became remotely green, and the later H-L2 request, decision, implementation, and activation crossed their ordered barriers. The sole real attempt then consumed at transport H0 before any body or EDF-header read. It cannot be rerun; the remaining 119 files, signals, targets, training, scoring, release, and claim upgrade remain unauthorized.

The parallel end-to-end plan is tracked in docs/COMMUNICATION_EEG_SCIENTIFIC_CLAIM_PROGRAM.md. It selects OpenNeuro ds003626-v2.1.2 for a future four-command EEG-versus-EOG-and-oral-EMG discovery test, then requires a separately qualified replication cohort, dropout-safe causal replay, source-only endpointing, and one consented unseen-person device stream. This Tier A plan made zero data, model, target, score, provider, stream, or hardware operations and does not change the Dreyer closeout or activate a new gate.

The generated-only live-session boundary is implemented, qualified, and closed in src/neurodecodekit/streaming/live_session.py and documented in docs/COMMUNICATION_LIVE_SESSION_G0_QUALIFICATION_CLOSEOUT.md. Its sole fictional qualification passed and is consumed; it has no scientific value. This is interface engineering only: it is not a live device run or neural result.

The next communication milestone is now prospectively frozen in docs/COMMUNICATION_EEG_SOURCE_IDENTITY_PREREGISTRATION.md. Its exact OpenNeuro query and deterministic selector would retain one common complete raw session for all ten participants, including the BDFs that preserve EEG, eye, and mouth channels, while parking above 10 GiB. This is registration only: no dataset-specific response was requested, and neither the generated implementation nor public metadata access is authorized yet.

The total incremental NeuroDecodeKit research-storage allowance is 20 GiB, but the communication raw-selection cap remains 10 GiB. The balance is safety room for bounded derivatives and atomic temporary output, not permission to expand the cohort. Approximately 73 GiB was free when this contract was recorded.

The additive generated canonicalizer is recorded in src/neurodecodekit/datasets/communication_eeg_source_identity.py and recorded in docs/COMMUNICATION_EEG_SOURCE_IDENTITY_IMPLEMENTATION.md. It has no network or real-data execution command. Exact implementation be6c44f passed both required jobs in CI 33038044421; the one official generated qualification then passed two deterministic replays and all 20 adversarial refusals in 0.072875 seconds at 22,069,248-byte peak RSS. It is consumed and must not be rerun. Read docs/COMMUNICATION_EEG_SOURCE_IDENTITY_GENERATED_QUALIFICATION_RESULT.md. Real metadata, payloads, signals, targets, models, scores, and claims remain closed.

Before any future communication discovery target is delivered, read the independent replication lock:

docs/COMMUNICATION_EEG_INDEPENDENT_REPLICATION_PREREGISTRATION.md;
registries/communication_eeg_independent_replication_contract.v0.json;
tests/test_communication_eeg_independent_replication_contract.py; and
the matching entry in registries/current_research_frontier.v0.json.

COMM-R0-REPLICATION-v0 freezes the unseen-person split, offline event-locked sample-causal feature family, nuisance residualizer, compact classifier, K - 1 class-shift EEG control ensemble, no-provider language-control arms, numerical gates, and one-score firewall before a discovery result exists. It also requires the exact independent source and its full or partial claim ceiling to become remotely green before discovery target delivery. SilentSpeech-EEG is not yet operationally qualified; TESSCCo and Kara One remain partial challenges. This is Tier A preregistration only and changes neither the active Dreyer gate nor any real-data authority.

The generated-only COMM-R0-G implementation now lives in src/neurodecodekit/experiments/comm_r0_generated.py; use python -m neurodecodekit.comm_r0_cli plan to inspect its bounded schedule. It separates the neural freeze from later provider-free language controls, requires an identity-bound complete freeze before synthetic target delivery, fits all three four-class cyclic derangements, records the offline trial oracle explicitly, and refuses to proxy missing partial-route sensors. One local development replay passed, but the official two-replay qualification remains inactive until a separate activation and proof closeout are remotely green. No real EEG or scientific claim is created by this implementation.

Exact generated result f64d4f3 passed both required jobs in CI 33038718183. Its proof-only closeout binds the three exact result artifacts without rerunning the qualification or reading ignored state. Exact closeout 4acd82b passed both required jobs in CI 33039371687, so COMM-L0 generated engineering is closed. Read the queued all-false packet in docs/COMMUNICATION_EEG_SOURCE_IDENTITY_METADATA_AUTHORIZATION_PACKET.md. COMM-L0-META proposes one future 361-byte OpenNeuro request, a 16 MiB response cap, zero BDF payload bytes, and one target-free private manifest. Amendment 1 adds generated boundary qualification, durable no-follow marker rules, exact decision binding across M1/M2/M3, and an 8 MiB disk envelope. It is not active, authorizes nothing now, and cannot displace DREYER-C5R-1-HL.

Exact request 12e4e9f passed both required jobs in CI 33062307015. Proof-only closeout 3e1322d passed both required jobs in CI 33063157757. The packet is now reviewable but remains queued and all-false; those green proofs do not authorize M1, a metadata request, or any EEG operation.

Before proposing an independent communication cohort, read docs/COMMUNICATION_EEG_REPLICATION_SOURCE_REFRESH_2026_08_27.md and its machine registry. The refresh found no public source yet qualified for the full EEG-beyond-eye-and-mouth replication claim. SilentSpeech-EEG is a strong watchlist source but lacks a verified stable public payload identity, complete manifest, dataset license, loader, and exact external-channel roster. The other reviewed sources provide only partial controls. No source was promoted, downloaded, or authorized by that Tier A pass.

Then read docs/COMMUNICATION_EEG_TRIANGULATED_REPLICATION_REFRESH_2026_08_27.md. The additive refresh records newly public OpenNeuro ds007591-v1.0.1 as a three-person full-sensor mechanistic bridge. It cannot enter the full-control replication router because n=3 < 10, and an exact one-sided participant sign-flip test cannot attain better than p=0.125. The actual router therefore has two mandatory keys: full-control ds003626 discovery attribution and an independently frozen TESSCCo partial transportability result. The bridge is nonrouting and cannot rescue either failure. ds007808-v1.0.0 remains outside the route because its approximately 1.575 TB, three-participant surface exceeds the participant and selected-raw gates. This refresh authorizes no metadata, payload, model, target, score, stream, or device operation. Exact refresh 85400fe passed both required jobs in CI 33128665324 and its proof closeout passed at 171d615 / CI 33129343388; both are on GitHub main. The proof binds three exact artifacts / 23,934 bytes and performs no neural-data or claim operation.

Then read docs/COMMUNICATION_EEG_TESSCCO_SOURCE_READINESS_2026_08_27.md. The paper-cited Kaggle dataset DOI resolves to dataset ID 9993149, but the landing returned HTTP 404 on 2026-08-27. TESSCCo remains the planned partial independent key but is not acquisition-ready, and no metadata API or EEG file was requested. Exact readiness result 3ed55ba passed both required jobs in CI 33130065908 and reached GitHub main. Read its proof-only closeout next; it binds three exact artifacts / 15,460 bytes and repeats no source or payload request. Exact closeout 752f99c passed both required jobs in CI 33130671683 and reached GitHub main.

Then read docs/COMMUNICATION_EEG_KARA_ONE_SOURCE_READINESS_2026_08_27.md. The official page exposes 14 participant archives totaling approximately 24 GB but no member-level remote manifest proving a complete all-eligible- participant EEG plus eye/face inventory within the frozen 10 GiB cap. Its associated paper reports 12 recruited and eight analyzed participants, so the exact target-free eligibility set is also unresolved. The registered router forbids participant dropping and parks an over-cap or unsplittable route. Kara One is therefore parked under the currently exposed source surface, no archive was requested, and no public partial replication route currently qualifies. Preserve the prospective synchronized EEG/EOG/bilateral oral-EMG fallback and do not create an acquisition packet from this Tier A result. Exact readiness result a9a9b23 passed both required jobs in CI 33131980667 and reached GitHub main. Read its proof-only closeout next; it binds three exact artifacts / 17,634 bytes and repeats no source or payload request.

Then read docs/COMMUNICATION_EEG_OBJECTIVE_EVIDENCE_AND_REPLICATION_DECISION_2026_08_27.md. Its machine-tested ledger records one narrow prior item of real directional task evidence and keeps every target communication requirement false. It separates ds003626 discovery, SilentSpeech-EEG full-control watchlist, TESSCCo/Kara One partial challenge, and Dreyer motor-method roles. This is routing evidence only: it does not promote a source, authorize acquisition, or change the sole active Tier C packet. Exact decision 9c3489c passed both required jobs in CI 33069830931.

COMM-G1 and its COMM-G2 proof successor were later implemented and each consumed at R0; read their closeouts near the end of this file. Do not rerun or repair either lane. Neither is real-data or scientific evidence, and neither authorizes OpenNeuro metadata, payload, signal, target, model evidence, score, stream, or device operations.

Historical BNCI-C3C5-1 Closed Lane

The generated G1 pipeline is complete, remotely proven, and consumed. Its synthetic result is engineering evidence only. The original direct-response Stage A attempt is also consumed at HTTP-302 refusal with zero payload body bytes.

The separately authorized signed-object recovery has now completed and is consumed. Read the current result first:

docs/BNCI_2014_001_STAGE_A_REDIRECT_RECOVERY_RESULT.md;
registries/bnci_2014_001_stage_a_redirect_recovery_result.v0.json; and
tests/test_bnci_2014_001_stage_a_redirect_recovery_result.py.

Control-plane commit 21cedd5 passed Base job 97691784130, Optional Neuro Readers job 97691784315, and CI 32811586786 before the one registered live invocation. The maintainer explicitly accepted its exact 779,873,919-byte transfer and irreversible-consumption risk.

The live recovery acquired and opaque-verified exactly 18 signed objects and 779,873,919 payload bytes in 113.066114 seconds at 42,893,312-byte peak RSS. Total network bytes were 780,275,745, and 92,111,536,128 free-disk bytes remained. The original 297-byte consumed marker stayed byte-identical.

MAT semantic opens, signal/event/target/label reads, model and training runs, prediction sets, target deliveries, and scores were all zero. The payload and private evidence stay Git-ignored. The recovery cannot be retried or rerun.

Exact Stage A result 96d7f0569a54b05f8031d2e3943658ef598e38a5 passed Base job 97700176631, Optional Neuro Readers job 97700176787, and CI 32814564120.

Read the generated-only Stage Q implementation and result next:

docs/BNCI_2014_001_STAGE_Q_IMPLEMENTATION.md;
registries/bnci_2014_001_stage_q_implementation.v0.json; and
registries/bnci_2014_001_stage_q_generated_result.v0.json.

The one generated qualification passed and is consumed. It processed 40,069,562 generated bytes, six task runs, and 288 trials in 3.657 seconds at 498,434,048-byte peak RSS. It opened no private or real path and ran no model, training, prediction, target delivery, or score.

Exact final implementation e5ca6a2 passed Base Python job 97723744136, Optional Neuro Readers job 97723744450, and CI 32822604745. Read docs/BNCI_2014_001_STAGE_Q_LIVE_ACTIVATION.md and its machine activation.

Initial activation commit e678095 failed Base job 97727563579 in CI 32823864546 because its tests assumed the prior implementation commit existed in the shallow Actions checkout. It had no effect and opened no private path.

Compatibility commit 52b681e passed Base job 97730713400, Optional job 97730713304, and CI 32824921855. The activation is rebound to that exact green commit.

Final activation 0e36993 passed Base job 97733856845, Optional job 97733856470, and CI 32825946085 before the one semantic invocation.

Read the current aggregate result next:

docs/BNCI_2014_001_STAGE_Q_RESULT.md;
registries/bnci_2014_001_stage_q_result.v0.json; and
tests/test_bnci_2014_001_stage_q_result.py.

Stage Q passed and is consumed. Exactly 18 MAT files, 108 task runs, and 5,184 trials validated with 22 EEG plus three EOG channels at 250 Hz. The executor created 72,666,213 private target-firewalled derivative bytes in 34.464444 seconds at 910,704,640-byte peak RSS. No held-out-T row entered a fold capability. Models, training, predictions, target deliveries, and scores were all zero.

Exact Stage Q result 9832ae5 passed Base job 97740061957, Optional Neuro Readers job 97740061602, and CI 32827957362. Stage P is now open only to a separately green implementation and activation.

Read the current implementation next:

docs/BNCI_2014_001_STAGE_P_T_IMPLEMENTATION.md;
registries/bnci_2014_001_stage_p_t_implementation.v0.json; and
tests/test_bnci_2014_001_stage_p_t_live.py.

The implementation adds the fold-scoped Stage Q loader, exact 9 x 6 x 48 completeness firewall, sequential fold-local model controller, aggregate hash-only prediction freeze, and separately activated Stage T target transport and scorer. It changes no frozen scientific model or gate.

Exact implementation 7ba4f7c passed Base Python job 97990455561, Optional Neuro Readers job 97990455765, and CI 32906104408.

Read the current activation next:

docs/BNCI_2014_001_STAGE_P_LIVE_ACTIVATION.md;
registries/bnci_2014_001_stage_p_implementation_activation.v0.json; and
tests/test_bnci_2014_001_stage_p_activation.py.

Stage P activation a49609b passed Base job 97992958552, Optional Neuro Readers job 97992958346, and CI 32906928931 before the one real execution.

Read the Stage P result next:

docs/BNCI_2014_001_STAGE_P_RESULT.md;
registries/bnci_2014_001_stage_p_prediction_freeze.v0.json; and
tests/test_bnci_2014_001_stage_p_prediction_freeze.py.

Stage P passed and is consumed: 468 fits, 495 inference/prediction sets, and 41,472 private target-blind prediction rows completed in 55.674016 seconds at 629,194,752-byte peak RSS. It emitted 11,298,505 private bytes and a 5,037-byte aggregate freeze. All nine source-only selectors chose E1. Held-out targets remained sealed and no score occurred. Repeated frozen-cap LBFGS convergence warnings were observed and recorded without a model change or rerun.

Exact prediction freeze 2517fd1 passed Base job 97996323987, Optional Neuro Readers job 97996324203, and CI 32908059166.

Read the Stage T activation next:

docs/BNCI_2014_001_STAGE_T_SCORING_ACTIVATION.md;
registries/bnci_2014_001_stage_t_scoring_activation.v0.json; and
tests/test_bnci_2014_001_stage_t_activation.py.

Stage T activation 06e29cb passed Base job 97998435433, Optional Neuro Readers job 97998435729, and CI 32908763353 before the sole score.

Read the Stage T result next:

docs/BNCI_2014_001_STAGE_T_RESULT.md;
registries/bnci_2014_001_stage_t_result.v0.json; and
tests/test_bnci_2014_001_stage_t_result.py.

The registered result is BNCIC3C5-R2: C3 and C5-partial both failed. Selected EEG reached 38.35% balanced accuracy versus 25% no-signal and 29.67% timing, but posterior EEG reached 39.24%, only 5/9 participants improved, and p=0.0664. EOG+EEG moved log loss in the desired direction but missed the frozen effect-size, consistency, and significance gates.

Exact R2 result af8dcc9 passed Base job 98001418110, Optional Neuro Readers job 98001417963, and CI 32909763799. The lane is closed and consumed with no tuning or rerun.

Immediate gate: artifact-only postmortem and a fresh independent replication design. Do not reopen any consumed target, prediction, or private artifact.

Scientific Goal And Sequence

The decisive bounded question is whether scalp EEG adds held-out information beyond the three recorded EOG channels, not merely whether a classifier can exploit protocol timing or eye movement. The fixed order remains:

G1 generated proof -> A opaque acquisition -> Q target-blind validation
                   -> P target-firewalled prediction freeze -> T one score

Stages Q, P, and T are complete and consumed at R2. This lane established a real held-out score but no registered scientific claim. A future independent pass can establish only participant-independent BNCI protocol-condition prediction and incremental EEG sensor information beyond recorded EOG under this protocol. It cannot establish thought or language decoding, movement intention, exclusive motor-cortex origin, live decoding, portable hardware, home use, or clinical utility.

Historical BNCI And UG1 Context

The BNCI source decision selected only 18 original MAT files in NEMAR nm000139 tag v1.0.2, totaling 779,873,919 bytes. The nine-fold protocol holds out one person completely. EEG-only C3 must beat no-signal, timing, derangement, temporal, and spatial controls. C5-partial requires EOG+EEG log loss to beat both recorded EOG alone and a size-matched EOG+deranged-EEG fusion at the participant level.

The older EEGMMIDB-UG1 Stage S-A2 invocation failed TLS verification before an HTTP response or EDF request. Its marker is consumed and immutable, its dependent lane is closed, and it must not be retried or substituted.

Historical UG1 Context

VR39P result closeout 474d49f53e1a46633b99b0ef2ff62b9abebe7a9c passed Base job 97317987437, Optional job 97317987655, and CI 32688563490. VR39P is consumed at aggregate R2: no cohort was frozen and Freewill/CIL1 is permanently parked. Do not touch its private or ignored state.

The independent next lane is EEGMMIDB-UG1. Read, in order:

docs/EEGMMIDB_UNSEEN_PARTICIPANT_GENERALIZATION_RESEARCH.md;
registries/eegmmidb_unseen_participant_generalization_research.v0.json;
docs/EEGMMIDB_UNSEEN_PARTICIPANT_GENERALIZATION_PREREGISTRATION.md;
registries/eegmmidb_unseen_participant_generalization_contract.v0.json;
docs/EEGMMIDB_UNSEEN_PARTICIPANT_GENERALIZATION_AUTHORIZATION_PACKET.md;
registries/eegmmidb_unseen_participant_generalization_authorization_request.v0.json; and
the three matching test modules.

UG1 freezes one source-pooled low-frequency model on S001-S015 and one final zero-calibration evaluation on completely fresh S016-S030 participants. The source execution LOSO gate must pass before any fresh payload request. The final execution gate must beat the stronger no-signal/timing control and every fixed derangement at the participant level. Imagery is an upgrade only.

The staged future envelope is 36 metadata requests, at most 36 new EDFs and 256 MiB payload transfer, 512 MiB incremental disk, 1 GiB peak RSS, one CPU thread, at most 300 fits and 640 prediction sets, one target delivery, one score, and zero retry or rerun. Every authority flag is false now.

Exact all-false request c642d90b646ff32c6d83e648f7d7779810605e11 passed Base job 97322606634, Optional Neuro job 97322606501, and CI 32690289547. Its proof-only closeout binds nine immutable artifacts totaling 56,516 bytes and changes no scope.

Exact proof closeout 9117b1db343be38248944c24e3d93cafc4058d98 passed Base job 97324487895, Optional Neuro job 97324488042, and CI 32690987778. UG1 was then identified as the sole active Tier C packet, and the maintainer's next exact message was continue.

Read the authorization decision document, machine decision, and matching test. The decision quotes the actual eight UTF-8 bytes, binds the unchanged 67,923- byte green packet, and authorizes only the registered G -> M -> S -> F -> T sequence after its own remote-green proof and a separate green narrowing amendment. The amendment is required because two target-free reviews found underspecified control gates, participant-ID isolation, prediction-freeze binding, exact completeness, causal state, and resource semantics.

Exact authorization decision 3e173f6dc61b1f6b32dcc9839aa74a67759b9b3f passed Base job 97333988408, Optional Neuro job 97333988474, and CI 32694496933.

Exact Amendment 1 commit 9ab8639b4d6fa28a30321687c82b185b13b1813a passed both required jobs in CI 32696449436 before implementation. Exact implementation commit da2be31a3ea4b7a438f86039c1d80b182e628ccf then passed Base job 97363993816, Optional Neuro Readers job 97363993465, and CI 32704970582 before the one Stage G invocation.

Read docs/EEGMMIDB_UNSEEN_PARTICIPANT_GENERALIZATION_IMPLEMENTATION.md and its two machine records. The sole generated/mock qualification passed all 17 case classes, 61 fits, 420 participant-condition prediction sets, exact causal and firewall replays, and every frozen refusal boundary. It used 25,975,920 generated input bytes in 443.246267 seconds at 276,856,832-byte peak RSS and emitted a 3,911-byte aggregate result. Real path, EDF, cache, raw-data, real target, and network counts were zero. The synthetic R4 route has no scientific claim value and qualification may not be repeated.

Stage G closeout 5cc3e0e9fd5739e8836ddb91252f18ca7849c824 passed Base job 97373297588, Optional job 97373297708, and CI 32708050897. Stage G is consumed and must not be repeated.

Read docs/EEGMMIDB_UNSEEN_PARTICIPANT_METADATA_AUTHORIZATION_PACKET.md, its machine request, and matching test. The all-false Stage M request proposes a generated/mock metadata-client qualification followed, only after a separate green implementation barrier, by one exact 36-request metadata invocation. The future real surface is HTTPS HEAD only, zero redirects, zero retries, zero response-body bytes, and no EDF parsing or download. It binds 20 exact Stage G artifacts totaling 305,662 bytes.

Exact Stage M request e2647d609a99997ac417dac5d8efb2dad61863a0 passed Base job 97376524550, Optional job 97376524804, and CI 32709110804.

Read docs/EEGMMIDB_UNSEEN_PARTICIPANT_METADATA_REQUEST_PROOF_CLOSEOUT.md, its machine proof, and matching test. The proof binds the three unchanged request artifacts totaling 20,189 bytes and performs only tracked-artifact and Git proof reads.

Exact proof closeout e9c11da94730e790aace3acc818e029abcbdc165 passed Base job 97379680508, Optional job 97379680751, and CI 32710175884. EEGMMIDB-UG1-M was then identified as the sole active Tier C packet, and the maintainer's next exact message was continue.

Read the Stage M authorization decision document, machine decision, and matching test. The decision quotes those eight UTF-8 bytes and binds six unchanged packet/proof artifacts totaling 29,320 bytes. It authorizes only the registered two-stage maximum after its own remote-green proof: generated/mock Stage M1 first, then one body-blind Stage M2 metadata invocation only after the exact M1 implementation and proof closeout are also remotely green.

Exact authorization decision 021bf8a1f2f12a8e7388a561535328cd0dc0dba2 passed Base job 97385926125, Optional job 97385926444, and CI 32712235191 before Stage M1 implementation.

Read docs/EEGMMIDB_UNSEEN_PARTICIPANT_METADATA_IMPLEMENTATION.md, its machine implementation record, the new isolated metadata module, sidecar CLI, and matching tests. The implementation freezes the exact HEAD-only client, canonical inventory and receipt, 20 generated cases, and resource/refusal boundaries. A new isolated metadata sidecar exposes plan and qualify but no live execute command; the proof-bound Stage G sidecar remains byte-identical.

Corrected implementation 1c68a775294e08013f6ef0780eb8901917699db0 passed Base job 97395810059, Optional Neuro Readers job 97395810337, and CI 32715529168. The sole Stage M1 qualification then passed all 20 generated cases with 297 mock HEAD requests, exact replay, and zero real request, body, EDF, payload, target, model, training, or score operations. It processed 8,354 generated fixture bytes in 0.017772541963495314 seconds at 33,341,440-byte peak RSS and emitted a 1,416-byte aggregate with SHA-256 9f91843e6a20f8794cf19105116b3bcf13a2a3deff496a3c44ff30ecbcfeafe3. The run is consumed and may not be repeated.

Stage M1 result 3b343d74604b0b0d0e0732f14cf593c8e057ebbf passed Base job 97399780409, Optional Neuro Readers job 97399780188, and CI 32716836238.

Read the Stage M1 proof-only closeout, machine proof, and matching test. It binds nine exact implementation/result artifacts totaling 82,278 bytes under canonical artifact-set SHA-256 8c586bb311c1afb09bfacaf611da150398c40121802dd9b2c4683036f9bbb3c1 without repeating qualification or performing a real operation.

Stage M1 proof-only closeout fd88d9d7ca9ffd3951eda295daa74a05ee4201a9 passed Base job 97402552811, Optional Neuro Readers job 97402552736, and CI 32717768039. The sole Stage M2 invocation then completed all 36 direct HEAD requests in 4.292717207921669 seconds at 38,141,952-byte peak RSS.

Read the Stage M2 result document, canonical inventory, receipt, machine result, and matching test. The inventory freezes 92,414,976 declared bytes: six missing source-fit files totaling 15,498,816 bytes and 30 fresh-final files totaling 76,916,160 bytes. ETag, Last-Modified, and Accept-Ranges were present for all 36 files. Redirect, retry, body, EDF, payload, target, model, training, and scoring counts were zero. The real metadata invocation is consumed.

Stage M2 result 818ef1f6384a03c3681d9d4ec01d6f88db4d2749 passed Base job 97405609600, Optional Neuro Readers job 97405609428, and CI 32718796222.

Read the Stage M2 proof-only closeout, machine proof, and matching test. It binds eight exact M1-proof/M2-result artifacts totaling 37,723 bytes under canonical artifact-set SHA-256 051651c6e31ecb9451d445b8ae5549e2e775da994b8c7913b9b9077cf3924388 without another network request or real operation.

Stage M2 proof-only closeout 00f795f860762a6e828b210cee52808e69571d53 passed Base job 97409224259, Optional Neuro Readers job 97409224710, and CI 32720013549. Stage M is fully closed and consumed.

Read docs/EEGMMIDB_UNSEEN_PARTICIPANT_SOURCE_ACQUISITION_AUTHORIZATION_PACKET.md, its machine request, and matching test. EEGMMIDB-UG1-SA is an all-false acquisition-only packet, distinct from the later Stage S source-LOSO and training gate. It binds eleven exact upstream artifacts totaling 116,444 bytes and freezes only one future checksum-manifest request plus six source EDF requests totaling 15,498,816 successful payload bytes.

Exact request 2085ea061d936bb18ef08e93fb7d3f874ef0f9d8 has now passed Base job 97417435948, Optional Neuro Readers job 97417435670, and CI 32722744301.

Read the request proof-only closeout, machine proof, and matching test. They bind the three unchanged request artifacts totaling 33,704 bytes under canonical artifact-set SHA-256 8f91ca0ae7e9a4ef47ce1a2ab7d71c9b24f805ae737587d02edf9b23b6deb4d0.

Proof closeout b0b4632ffbdaca10c1e4bbc93ad26ebd8e1368ca passed Base job 97422237667, Optional Neuro Readers job 97422237272, and CI 32724357118. Codex named Stage S-A as the sole active Tier C packet; the maintainer's next exact message was continue.

Read the authorization decision, machine decision, and matching test. They quote those exact eight UTF-8 bytes and bind six unchanged request/proof artifacts totaling 44,910 bytes. Exact decision 1b5c9195f384e5867f18131aa7d669f7c9cd0e2b passed Base job 97426157639, Optional Neuro Readers job 97426157381, and CI 32725633524 before Stage S-A1 implementation.

Read the Stage S-A1 implementation document, machine registry, isolated source-acquisition module, sidecar CLI, and matching tests. The implementation freezes exact checksum parsing and seven-request order, 1 MiB streaming, six opaque local hash passes, a synced pre-request consumed marker, complete- bundle-only no-replace promotion, narrow cleanup, resource caps, and 27 generated adversarial cases. The CLI exposes only plan and qualify; no live command or generic URL/path override exists. Corrected implementation 37808bef8c59bc862345f342fd932aa04373b3fd passed Base job 97441967842, Optional Neuro Readers job 97441968304, and CI 32730673153 before the one registered generated qualification.

Read the Stage S-A1 result document, machine result, and matching test. All 27 cases passed with 56 mock requests, three complete generated bundles, 67,199,028 body bytes read, 18 post-write passes, 0.4758401670260355 seconds runtime, 71,696,384-byte peak RSS, and zero retention. The raw generic nested counter block incorrectly reports both mock subtypes as zero; the immutable case flow exactly reconciles 21 checksum plus 35 EDF mock requests. The raw aggregate is hash-bound, the discrepancy is explicit, and the consumed qualification must not be rerun.

Exact result e2965cfc90ac73a0371689333d0bad67de2634fd passed Base job 97450079515, Optional Neuro Readers job 97450079739, and CI 32733249548.

Read the Stage S-A1 proof-only closeout, machine proof, and matching test. The pending closeout binds nine exact artifacts totaling 130,558 bytes under canonical SHA-256 859988a3e658c8f8872d9ffe82151f17bc475c94e754e817dd3c3dd698fd7ad5 without repeating qualification or performing a real operation.

Exact proof-only closeout b3902cf50bef478055255570d1b78813207fb8d1 passed Base job 97456050452, Optional Neuro Readers job 97456050604, and CI 32735141922.

Read the Stage S-A1 green-proof activation document, machine activation, updated proof registry, and matching tests. The activation fills only the frozen green-proof evidence fields, performs no real operation, and remains ineffective until its own commit passes both CI jobs.

Immediate gate: commit, push, and remotely green the exact activation. Stage S-A2 network access remains closed now. No local real payload, EDF semantic content, target, model, training, score, release, or claim action is open, and the 54 retained source EDFs plus all 30 fresh-final EDFs remain untouched.

Previous Gate: MARC2-VR39P Aggregate R2 Closeout

VR38A proof-only closeout a599adf3e0320ad420e1c2f5647a0432e645c246 passed Base job 97278761357, Optional job 97278761303, and CI 32673882729, remotely closing the generated selection-sufficiency repair. Read docs/MARC_2_SELECTION_SUFFICIENCY_PRIVATE_COHORT_FREEZE_AUTHORIZATION_PACKET.md, its machine request, its request test, docs/MARC_2_SELECTION_SUFFICIENCY_PRIVATE_COHORT_FREEZE_REQUEST_PROOF_CLOSEOUT.md, and registries/marc2_selection_sufficiency_private_cohort_freeze_request_proof.v0.json.

Exact all-false request 6c805a817fa44375b7b0e120abcb2c748c78ca07 passed Base Python job 97283907786, Optional Neuro Readers job 97283907932, and CI 32675925646. Its three exact request artifacts total 45,742 bytes and remain unchanged.

The remotely green closeout honestly records one repository-root listing that exposed only the .codex_work root entry's metadata. No metadata value is retained, no descendant or content was accessed, and the deviation neither authorizes nor consumes a stage.

Exact request proof e581fe99d97e91e5af07e211dd75caa22a10d098 passed Base job 97288870035, Optional job 97288870180, and CI 32677755105. After VR39P was identified as the sole active Tier C packet, the maintainer's next exact message was continue. Decision dbde5f84b3fac0ac0b23208afd56e00d678aff00 passed Base job 97298894039, Optional job 97298894171, and CI 32681510484 before Stage 1.

The independent wrapper then passed its sole generated qualification: 168 paths, 168 VR33A calls, 504 provider calls, 336 sleeper calls, 84 source opens, 84 VR38A calls, 64 temporary cohort writes, exact R1/R2 64/104, 268 direct refusals, 12 named critical witness classes, exact replay, zero mutation, and zero retained output. Runtime was 12.816678500035778 seconds at 67,158,016-byte peak RSS. Generated input was 37,159,376 bytes and peak live output was 313,627 bytes.

Only after all later proof barriers may one target-free 418,755-byte source read either write a private HMAC-committed cohort or permanently park Freewill/CIL1.

The public surface has only R1 cohort frozen and R2 lane parked. It exposes no count, identity, topology, row, path, storage total, runtime, RSS, readiness state, failure stage, raw hash, or deterministic selection hash. The private cohort preserves sessions 1/2, runs 1-3, exact split roles, source rows, distinct raw/canonical source hashes, and compressed/uncompressed storage feasibility under the 10 GiB project ceiling.

Exact Stage 1 4d48cb38822e3e5a819ce1fef0188069ca6bd9ac passed Base job 97310285688, Optional job 97310285728, and CI 32685719113. The separate proof-only closeout now binds that immutable implementation and all eight exact implementation Git blobs. It repeats no qualification, touches no private state, and remains ineffective before its own remote green proof.

Exact proof-only closeout cec5fe87a6ddc122366e0db32e2c5147bae47c81 passed Base job 97313196679, Optional job 97313196627, and CI 32686765350. Activation c74255526a9265e52919672e8c6088022b77ef1d then passed Base job 97314783863, Optional job 97314783869, and CI 32687368688 before the sole Stage 2 invocation.

The executor returned aggregate MARC2VR39P-R2 with no cohort commitment. R2 means only that no cohort was frozen and the Freewill/CIL1 lane is permanently parked. Readiness, failure stage, source operations, topology, identity, storage, runtime, RSS, and output measurements are unavailable by contract. No private output or source was inspected after consumption.

Immediate gate: commit, push, and green the aggregate-only R2 result closeout. Then move to an independent preregistered scientific lane; VR39P has no retry or successor.

Recent Structural History

VR36P requests a generated/mock wrapper followed, after every proof barrier, by one fixed target-free structural read and one VR35A call. A successful G1 or G2 route may freeze one 16-subject, 96-bundle, 384-row private cohort. It does not open archive members, neural data, targets, models, or scores.

Exact all-false request 8ec87ced3c0072fec62328a9635eb9774e13e605 passed Base job 97211815865, Optional job 97211815879, and CI 32646648532. Request proof closeout 2813d60707f9fd97acbfa35cd57504d6b3db86c5 passed Base job 97213774526, Optional job 97213774452, and CI 32647453505.

After VR36P was identified as the sole active Tier C packet, the maintainer's next exact message was coninue. The authorization decision preserves those seven UTF-8 bytes and binds only the unchanged packet. Exact decision fd08dd6ee40b16d3b4f4312601fed3370b7e2ca5 passed Base job 97215989173, Optional job 97215989332, and CI 32648347577 before Stage 1.

The independent fixed-path wrapper then passed its sole 40-path generated qualification. VR36P R1-R5 each appeared four times and R6 twenty times; 40 VR33A calls, 120 provider calls, 80 sleeper calls, 20 VR35A calls, eight temporary cohort writes, 111 direct refusals, exact replay, and zero source mutation passed. Runtime was 1.785233375034295 seconds at 35,340,288-byte peak RSS, with zero retained output and zero private operation.

Exact implementation 8179f6fd4acb721ef25b023e02ac9160789f9d49 passed Base job 97220389999, Optional job 97220389862, and CI 32650171033. Proof-only closeout d4074081b86e6b6247f91150daa1e3253f6e2bd9 passed Base job 97222491265, Optional job 97222491354, and CI 32651006809 before the sole invocation.

The executor returned MARC2VR36P-R3 after exactly three readiness samples, two sleeps, one 418,755-byte target-free structural open/parse, one VR33A call, and one VR35A call. Runtime was 10.088516000076197 seconds at 30,670,848-byte peak RSS. No cohort was frozen.

VR36P R3 means only that exact-task projection still leaves the eligible total above 195. Mixed-task contamination is therefore not the sole explanation for the aggregate surplus, but the exact count, difference, distribution, identity, participant, and cause remain unavailable. VR36P is consumed with no rerun or private reinspection.

VR37A registration a677e7abd2b89e92bb7bcc3f823a3493c6a32ad0 passed Base job 97226913287, Optional job 97226913421, and CI 32652807264 before implementation. Its sole generated qualification passed 24 paths with G1 and R1-R5 each appearing four times, exactly 24 unchanged VR35A calls, 87 direct refusals, exact replay, and zero source mutation or retention.

The pass processed 10,606,968 generated bytes in 2.4954585829982534 seconds at 49,266,688-byte peak RSS. It distinguishes baseline, single-cell contiguous and noncontiguous extensions, multi-cell pure surplus, mixed surplus/deficit with positive net, and structural-firewall refusal on generated fixtures only. It does not identify the consumed private topology.

Exact implementation f698a10d2649bd65f7b819531cbf3d89cd3f0c0a passed Base job 97230546153, Optional job 97230546048, and CI 32654295221. The proof-only closeout now binds the preproof registry hashes and every exact implementation Git blob. Qualification was not repeated and private operation count remains zero.

Proof-only closeout 4287a860ae11b47a9ccb50090ae62e6e4be2b59a passed Base job 97232723710, Optional job 97232723664, and CI 32655208775. VR37A is remotely closed as generated engineering.

VR37P is now an all-false Tier C request for a future two-stage sequence: generated/mock fixed-path qualification first, then only after its own green implementation and proof closeout one route-only private topology read. The future output may expose only R1-R7, never the exact count, difference, cell, identity, run, row, or path. It freezes no cohort.

Exact request e89908adf508ca3c858b4ee0509c45fd8d29f866 passed Base job 97234876608, Optional job 97234876537, and CI 32656051436. Its proof closeout binds the 5,743-byte packet, 15,681-byte machine request, and 7,715-byte request test by SHA-256 and Git blob without changing scope.

Proof closeout 7de128053620514d067dbc1c99318a2b3bdb69e1 passed Base job 97236607433, Optional job 97236607490, and CI 32656773778. VR37P is therefore remotely proven, but no packet-bound decision was recorded and it remains unexecuted. Because every VR37P route freezes zero cohorts, executing it would not unlock FW2 or CIL1.

The generated-only MARC2-VR38A selection-sufficiency repair then passed its sole qualification as summarized above. It preserved exact source, task, taxonomy, companion, rank, split, and storage checks while removing only the unnecessary global exact-195 requirement. Its implementation and proof-only closeout are remotely green. No real cohort or neural payload has been opened.

Consumed Boundary: MARC2-VR34P Protocol-Conforming R2

Read docs/MARC_2_EXACT_COUNT_PRIVATE_CONFIRMATION_RESULT.md and registries/marc2_exact_count_private_confirmation_private_result.v0.json. Proof-only closeout c863fd4b5bb9c866d6bc5683cb74c9fbeab6d8d8 passed Base Python job 97198447774, Optional Neuro Readers job 97198447418, and CI 32641201220 before the sole invocation.

The executor collected exactly three readiness samples with two fixed sleeps, opened and strict-parsed exactly 418,755 target-free structural bytes once, called VR33A once, called VR31A once, and returned MARC2VR34P-R2. Runtime was 10.147974541061558 seconds at 26,935,296-byte peak RSS under one thread, worker, and numerical job. Network and new-payload bytes were zero.

R2 means only that the filtered eligible total is above 195. The exact count, difference magnitude, private value, participant, selection, and cohort remain unavailable. No cohort was frozen. Unlike consumed VR32P, this invocation matched the registered three-sample, two-sleep readiness protocol exactly.

VR34P cannot be retried, repaired, or reinspected. The next safe task is an artifact-only task-aware eligibility repair, now frozen as VR35A. Archive, neural, target, model, score, FW2/CIL1, device, release, and scientific-claim surfaces remain closed.

Consumed Boundary: MARC2-VR32P R2 With A Protocol Deviation

Read docs/MARC_2_ELIGIBLE_TOTAL_DIRECTION_PRIVATE_DISCRIMINATOR_RESULT.md and registries/marc2_eligible_total_direction_private_discriminator_private_result.v0.json. Exact proof-only closeout 5aec8a15f5ee6fa3c6ca9cefcfb4fbfead9dd72f passed Base Python job 97177262321, Optional Neuro Readers job 97177262387, and CI 32632497701 before the sole invocation.

The executor opened and strict-parsed exactly 418,755 target-free structural bytes once, called VR31A once, and returned MARC2VR32P-R2. Runtime was 20.141367041040212 seconds at 30,130,176-byte peak RSS under one thread, worker, and numerical job. Network and new-payload bytes were zero.

R2 means only that the filtered eligible total is above 195. The exact total, difference magnitude, participant, selection, and cohort remain unavailable, and no cohort was frozen.

The returned report also records five readiness samples against the packet's registered three. The invocation is therefore consumed with a readiness- sampling protocol deviation and is not a fully protocol-conforming registered result. It cannot be retried or repaired.

The exact-count readiness repair is remotely closed as VR33A. The next all-false request is VR34P; it performs no private operation during packet preparation. Do not inspect VR32P private or consumed state. Archive, neural, target, model, score, FW2/CIL1, device, release, and scientific-claim surfaces remain closed.

Generated Evidence: MARC2-VR31A Remotely Closed

Read docs/MARC_2_R1_ELIGIBLE_TOTAL_DIRECTION_DISCRIMINATOR_IMPLEMENTATION.md and both matching machine records. Registration eeab6785b8eadc6d65199fa1ac519173f9c160c7 passed Base Python job 97163443088, Optional Neuro Readers job 97163443152, and CI 32626878097 before implementation.

The one measured generated-only qualification returned MARC2VR31A-G1. All 32 paths passed with 32 unchanged VR29A calls, eight R1 direction comparisons, one immutable threshold predicate, and exact G1/G2/R1/R2/R3 counts 4/4/4/4/16. It also passed 78 direct refusals, exact replay, order-invariant route distribution, source immutability, zero observed-count retention, and zero retained output.

The pass processed 14,137,216 generated bytes in 2.8035786249674857 seconds at 39,174,144-byte peak RSS. The canonical aggregate report was 2,957 bytes. Every private, ignored-path, readiness, archive, neural, target, model, score, network, FW2/CIL1, hardware, other-project, and claim counter was zero.

Exact implementation 3553fa4f7d3d7e4b5e3813ac3f6219cf6ba759ab passed Base Python job 97165896566, Optional Neuro Readers job 97165896610, and CI 32627856478. The proof-only closeout now binds both preproof registries and eight exact implementation Git blobs. It repeats no qualification and performs zero private operation.

The exact proof-only closeout above is remotely green. VR31A is closed as generated engineering. It has no private executor and establishes no real direction, cohort, neural effect, or decoding result.

Current Evidence Boundary: MARC2-VR30P Consumed At R1

Read docs/MARC_2_INVENTORY_DISTRIBUTION_PRIVATE_DISCRIMINATOR_RESULT.md and registries/marc2_inventory_distribution_private_discriminator_private_result.v0.json first. Proof-only closeout 7ce2c5e7f7dde15dfe1dfafa35058613ae09b016 passed Base Python job 97159363190, Optional Neuro Readers job 97159363046, and CI 32625197776 before the sole invocation.

The executor collected three fresh readiness samples, opened and strict-parsed exactly 418,755 target-free structural bytes once, called VR29A once, called unchanged VR25A once, called the unchanged VR2 eligible filter once, and consumed at MARC2VR30P-R1. Runtime was 10.091467417078093 seconds at 30,752,768-byte peak RSS under one CPU thread, one worker, and one numerical job. Network and new-payload bytes were zero.

R1 means the filtered eligible total differs from the registered public total of 195. It does not reveal the observed total, direction, magnitude, failed private value, row, path, identity, participant, selection, or cohort. Because the route is ordered, it also does not establish whether the later participant- session distribution check would pass. No cohort was frozen.

Exact combined readiness, marker, and report bytes were not returned and were not recovered through private inspection; the 1,246-byte canonical aggregate return is the only reconstructed output measurement. Post-execution free disk was 113,453,273,088 bytes.

VR30P is consumed with no retry, rerun, resume, repair, fallback, substitution, cleanup, amendment, output inspection, or private reinspection. Do not touch its private source, readiness state, consumed marker, or output root. The next safe task is an artifact-only and generated-only direction discriminator for below- versus above-expected eligible totals. Any new private read remains a new Tier C packet and decision. FW2/CIL1, archive and neural payload, targets, models, scores, devices, streams, releases, and scientific claims remain closed.

Generated Evidence: MARC2-VR29A Remotely Closed

Registration fcd088cc2eef6556f36ed596c6d9bb6c7ee9d7c3 passed CI 32618866986 before implementation. The one measured generated-only qualification returned MARC2VR29A-G1: all 32 paths passed with 32 unchanged VR25A calls, 16 additional unchanged VR2 eligible-filter calls, G1/G2/R1/R2/R3 counts 4/4/8/8/8, 77 direct refusals, exact replay, order- invariant route distribution, source immutability, and zero retention.

The pass processed 14,137,216 generated bytes in 2.2789369999663904 seconds at 37,371,904-byte peak RSS and retained zero. Exact implementation 2e73c9176d243b5deccbf8416bb59fdf053ba762 passed CI 32620018855; exact proof-only closeout 80badb9c1410c1661403aae966b1ea31fa0a45f1 passed CI 32620685817. This proves only generated discrimination of eligible-total versus participant-session distribution arithmetic. It does not identify the consumed private R1 subclass, establish a cohort, access neural data, or establish decoding performance.

Current Evidence Boundary: MARC2-VR28P Consumed At R1

Read docs/MARC_2_INVENTORY_TAXONOMY_PRIVATE_DISCRIMINATOR_RESULT.md and registries/marc2_inventory_taxonomy_private_discriminator_private_result.v0.json first. Proof-only closeout 96bff687013dcbfb507455b5f8c045977bc84fe8 passed Base Python job 97139815087, Optional Neuro Readers job 97139815147, and CI 32617240661 before the sole invocation.

The executor collected three fresh readiness samples, opened and strict-parsed exactly 418,755 target-free structural bytes once, called VR25A once, called the frozen VR27A map once, and consumed at MARC2VR28P-R1. Runtime was 10.065900957910344 seconds at 30,654,464-byte peak RSS under one CPU thread, one worker, and one numerical job. Network and new-payload bytes were zero.

R1 excludes unknown-participant taxonomy and localizes the remaining blocker to eligible-inventory or participant-session distribution validation. It does not identify the failed predicate, count, direction, row, path, identity, participant, selection, or cohort. Exact combined readiness, marker, and report bytes were not returned and were not recovered through forbidden post-result inspection; the 1,183-byte canonical aggregate return is the only reconstructed output measurement.

VR28P is consumed with no retry, rerun, resume, repair, fallback, substitution, cleanup, amendment, output inspection, or private reinspection. Do not touch its private source, readiness state, consumed marker, or output root. The next safe task is an artifact-only and generated-only R1 decomposition of eligible-inventory arithmetic versus participant-session distribution arithmetic. Any later private read remains a new Tier C packet and decision. FW2/CIL1, archive/neural payload, targets, models, scores, devices, streams, releases, and scientific claims remain closed.

VR24P result a873f1a passed Base job 97103071419, Optional Neuro Readers job 97103071353, and CI 32602610854. The exact one-shot result remains consumed at MARC2VR24P-R2; do not retry, reopen, list, hash, or inspect its private state.

Read docs/MARC_2_COMPLETE_BUNDLE_ARITHMETIC_RESEARCH.md, docs/MARC_2_SELECTION_BOUNDARY_FIREWALL_PREREGISTRATION.md, and registries/marc2_selection_boundary_firewall_contract.v0.json. VR25A separates a full-source compatibility count from the stronger invariants that actually protect a study cohort. It keeps exact 1,227-row and 1,025/202 entry- kind validation, validates every row and complete companion group, requires the exact 195 eligible bundles and participant/session distribution, then quarantines known ineligible groups before the unchanged target-free prefix selection.

The frozen generated matrix has ten cases and 40 paths. Five accepted controls must produce one identical 16-subject, 96-bundle, 384-member selection even when only known non-selected bundle counts vary. Added, removed, or redistributed eligible bundles, unknown participants, and incomplete companion sets must refuse. A count mismatch may expose only a boolean warning, never the observed count or difference.

Registration ad8be21 passed Base job 97105227375, Optional Neuro Readers job 97105227440, and CI 32603540967 before implementation. The additive standard-library module then passed the one registered generated qualification: 40/40 matrix paths, 20 accepted paths with one exact selection identity, 20 expected eligible/unknown/incomplete refusals, 72 direct refusals, and 40 source-immutability checks.

The pass used 24,580,902 generated input bytes, 3.250688 seconds, 48,185,344- byte peak RSS, and zero retained output. It has only plan, qualify, and inspect; there is no private executor or path argument.

Exact implementation 891245d passed Base job 97108121455, Optional Neuro Readers job 97108121321, and CI 32604761988. The proof-only closeout binds the two preproof registries and seven exact Git blobs without repeating the qualification or touching private state.

Exact proof closeout 378e863 passed Base job 97109778233, Optional Neuro Readers job 97109778216, and CI 32605475758 without repeating qualification or touching private state.

Read docs/MARC_2_SELECTION_BOUNDARY_PRIVATE_CONFIRMATION_AUTHORIZATION_PACKET.md and its machine request. VR26P is an all-false request for a future generated fixed wrapper and, only after separate proof barriers and a fresh Tier C decision, one 418,755-byte target-free structural read. A successful G1 or G2 route would freeze the same exact 16-subject, 96-bundle, 384-member cohort; only the public compatibility boolean differs. The observed complete-bundle count and direction remain unavailable and may not be inferred or published.

Exact request 00db825 passed Base job 97112059257, Optional Neuro Readers job 97112059152, and CI 32606451461. The proof-only request closeout binds the unchanged 10,278-byte packet, 17,193-byte request, and 12,705-byte request test with exact hashes and Git blobs. It performs zero implementation or private operation.

Exact request proof closeout efd779a passed Base job 97114173204, Optional Neuro Readers job 97114173360, and CI 32607272954. Codex then named VR26P as the sole active Tier C packet and restated its immutable two-stage scope. The maintainer's next exact message was the eight UTF-8 bytes continue.

Read docs/MARC_2_SELECTION_BOUNDARY_PRIVATE_CONFIRMATION_AUTHORIZATION_DECISION.md and its machine decision. They bind only the unchanged green packet and have delayed effect. Exact decision 8b9ce24 passed Base job 97116976065, Optional Neuro Readers job 97116976016, and CI 32608318342 before Stage 1.

Read docs/MARC_2_SELECTION_BOUNDARY_PRIVATE_CONFIRMATION_IMPLEMENTATION.md and its machine record. The additive standard-library wrapper passed 40/40 generated paths: 20 accepted, 20 expected refusals, 40 exact VR25A calls, 106 direct refusals, and deterministic replay. All accepted paths produced the same 16-subject, 96-bundle, 384-member selection. Four compatibility-true and 16 compatibility-false accepted paths confirmed that the warning boolean does not change cohort identity.

The pass used 17,672,740 generated input bytes, 2.444167 seconds, 34,701,312- byte peak RSS, and 223,419 peak incremental output bytes, with zero retained output and zero real/private path operations. Exact implementation d5c5abd passed Base job 97119373912, Optional Neuro Readers job 97119373779, and CI 32609236180.

Read docs/MARC_2_SELECTION_BOUNDARY_PRIVATE_CONFIRMATION_PROOF_CLOSEOUT.md. It binds the exact remote-green implementation, 7,417-byte preproof registry, canonical artifact-set hash, and four implementation Git blobs without repeating qualification or touching readiness or private state. Exact closeout bb6c52b passed Base job 97120986292, Optional Neuro Readers job 97120986207, and CI 32609855945 before the sole invocation.

Read docs/MARC_2_SELECTION_BOUNDARY_PRIVATE_CONFIRMATION_RESULT.md and its machine result. The executor collected three fresh readiness samples, opened and strict-parsed exactly 418,755 target-free structural bytes once, called VR25A once, and consumed at MARC2VR26P-R5 with no cohort freeze. Runtime was 0.0482423750218004 seconds at 29,622,272-byte peak RSS; combined output was 2,923 bytes, and all neural, target, model, score, network, and claim counters were zero.

R5 means only participant taxonomy or exact eligible inventory refused. It does not distinguish eligible count/distribution drift from an unknown recognized participant, and no private predicate, value, count, identity, selection, or cohort was retained. Do not retry or inspect VR26P private state. The next safe task is an artifact-only and generated-only R5 decomposition. FW2/CIL1 and all neural or scientific work remain closed.

VR26P result 878148a passed Base Python job 97122530294, Optional Neuro Readers job 97122530346, and CI 32610456792. Read docs/MARC_2_R5_INVENTORY_TAXONOMY_DISCRIMINATOR_PREREGISTRATION.md and its machine contract. VR27A freezes five generated cases across two orders and two replays: one exact control, three independent eligible-inventory mutations, and one unknown-participant mutation. The twenty paths must call unchanged VR25A exactly once each and return G1=4, R1=12, and R2=4.

Immediate gate: commit, push, and obtain both green CI jobs for the exact VR27A registration before implementation. The lane has no private executor and cannot identify the consumed private branch. Any later private read requires a separate all-false Tier C packet and fresh packet-bound decision.

Registration 47ceba3 passed Base Python job 97124216923, Optional Neuro Readers job 97124216871, and CI 32611101033 before implementation. Read docs/MARC_2_R5_INVENTORY_TAXONOMY_DISCRIMINATOR_IMPLEMENTATION.md. The generated adapter passed all 20 paths with exact route counts G1=4, R1=12, and R2=4, 20 exact VR25A calls, 57 direct refusals, exact replay, and zero source mutations. Runtime was 1.120283 seconds at 33,439,744-byte peak RSS over 8,836,136 generated input bytes, with 2,561 aggregate bytes and zero retention.

Immediate gate: commit, push, and green the exact implementation and result, then add and separately green a proof-only closeout. Do not prepare a private packet before both barriers. VR27A has no private executor, resolves no real branch, and establishes no neural or decoding result.

Exact implementation 3f74be3 passed Base Python job 97126099642, Optional Neuro Readers job 97126099573, and CI 32611864949. Read docs/MARC_2_R5_INVENTORY_TAXONOMY_DISCRIMINATOR_PROOF_CLOSEOUT.md. It binds the two preproof registries and seven exact implementation-commit Git blobs without repeating qualification or touching private state.

Exact proof-only closeout f6b5dbf passed Base Python job 97127512656, Optional Neuro Readers job 97127512634, and CI 32612454458. VR27A is closed as remotely proven generated engineering. It distinguishes eligible-inventory or participant-session distribution drift from unknown- participant taxonomy on generated witnesses only.

Read docs/MARC_2_INVENTORY_TAXONOMY_PRIVATE_DISCRIMINATOR_AUTHORIZATION_PACKET.md and its machine request. VR28P is an all-false two-stage request: a future generated fixed-path wrapper, then only after every proof barrier and a fresh Tier C decision one 418,755-byte target-free structural read, one VR25A call, and one VR27A map call. It may retain only aggregate R1 inventory/distribution or R2 unknown-taxonomy evidence. It cannot retain private counts, directions, rows, paths, identities, participants, selections, or a cohort.

Exact request 4e5895f passed Base Python job 97130420447, Optional Neuro Readers job 97130420507, and CI 32613575234. Read docs/MARC_2_INVENTORY_TAXONOMY_PRIVATE_DISCRIMINATOR_REQUEST_PROOF_CLOSEOUT.md and its machine proof. The closeout binds the unchanged 8,318-byte packet, 14,289-byte request, and 7,580-byte test by hash and Git blob. It changes no scope and performs zero implementation or private operation.

Exact proof-only closeout 87d4b30 passed Base Python job 97131943291, Optional Neuro Readers job 97131943377, and CI 32614149622. Codex then identified VR28P as the sole active Tier C packet. The maintainer's next exact message was the eight UTF-8 bytes continue.

Read docs/MARC_2_INVENTORY_TAXONOMY_PRIVATE_DISCRIMINATOR_AUTHORIZATION_DECISION.md and its machine decision. They bind only the unchanged green packet by reference and have delayed effect.

Exact decision 718c3de passed Base Python job 97133595196, Optional Neuro Readers job 97133595235, and CI 32614796767 before Stage 1. Read docs/MARC_2_INVENTORY_TAXONOMY_PRIVATE_DISCRIMINATOR_IMPLEMENTATION.md and the matching implementation and result registries. The generated fixed-path wrapper passed all 20 cases with 20 source opens, 20 exact VR25A calls, 20 VR27A map calls, expected G1/R1/R2 counts 4/12/4, deterministic replay, marker-before-open ordering, and 110 direct refusals.

The pass used 8,836,136 generated input bytes, 1.097564 seconds, 34,177,024- byte peak RSS, 785 peak incremental output bytes, and zero retained output. Every private, consumed-state, archive, neural, target, model, score, network, device, FW2/CIL1, other-project, and claim counter remained zero.

Initial implementation 06174e1 passed Base Python job 97137205238, Optional Neuro Readers job 97137205346, and CI 32616187929. A generated- only proof-transition test was then hardened without changing the module or measured result. Exact hardened implementation 6d3b770 passed Base Python job 97138335047, Optional Neuro Readers job 97138335116, and CI 32616632414.

Read docs/MARC_2_INVENTORY_TAXONOMY_PRIVATE_DISCRIMINATOR_PROOF_CLOSEOUT.md. It binds the exact preproof implementation and result registries plus five implementation Git blobs and repeats no qualification or private operation.

Exact proof-only closeout 96bff687013dcbfb507455b5f8c045977bc84fe8 passed Base Python job 97139815087, Optional Neuro Readers job 97139815147, and CI 32617240661 before the sole Stage 2 invocation. That invocation is now consumed at R1 as recorded at the top of this file. Do not retry or reinspect it.

Historical Frontier: MARC2-VR24P Consumed At R2

Read docs/MARC_2_R5_PRIVATE_DISCRIMINATOR_RESULT.md and its machine registry first. Proof-only closeout ea39319 passed both jobs in CI 32594732196 before the sole registered invocation. The executor opened and strict-parsed exactly 418,755 target-free structural bytes once, called VR20A once, applied the VR21A route map once, and consumed at MARC2VR22P-R4 without freezing a cohort. Runtime was 0.056165 seconds at 30,900,224-byte peak RSS; combined output was 2,872 bytes; network and new payload bytes were zero.

R4 narrows the blocker to VR20A F06. The two public possibilities are source-kind/run-bundle-total arithmetic and participant-taxonomy/eligibility arithmetic. F07 selection, split, rank, and reservation arithmetic is excluded for this execution. The exact private predicate and value were deliberately not retained. Do not retry or inspect consumed VR22P private state.

Registration cee91b0 passed both jobs in CI 32596045581 before implementation. The additive standard-library VR23A adapter then passed all 24 frozen generated paths. G1 and R1-R5 each appeared four times across six cases, two orders, and two replays; unchanged VR20A was called 24 times; diagnostic/upstream route disagreements and source mutations were zero; and 62 direct refusals passed. Static proof bound two VR20A F06 wrapper sites and seven VR2 helper reasons while proving two defensive branches non-independent.

The pass processed 10,603,766 generated bytes in 2.412034 seconds at 39,944,192-byte peak RSS and retained zero output. Exact implementation 9e1b121 passed both jobs in CI 32596999769; proof-only closeout 7d6b6c5 then passed both jobs in CI 32597604907 without repeating qualification or touching private state.

The all-false MARC2-VR24P request now binds 16 tracked inputs totaling 224,299 bytes. It proposes a future generated fixed-path wrapper and, only after separate green barriers and a fresh decision, one possible 418,755-byte target-free structural read that retains one aggregate R1-R5 class. It does not request a cohort freeze. Every authority flag is false and every operation counter is zero. Exact request b7fca40 passed both jobs in CI 32598316430. The request proof closeout binds the three unchanged request artifacts totaling 29,829 bytes and performs zero private or implementation operations. Exact proof closeout c53d4cf passed both jobs in CI 32598893728. Codex then identified VR24P as the sole active Tier C gate, and the maintainer's next exact message was continue.

The packet-bound decision preserves those eight UTF-8 bytes and authorizes only the unchanged two-stage sequence by reference. Exact decision 53bc114 passed both jobs in CI 32599996090 before Stage 1.

The generated fixed-path wrapper passes 24/24 paths: G1 and R1-R5 each four times, 24 exact VR23A calls, 24 nested unchanged VR20A calls, deterministic replay, and 130 direct refusals. It processed 10,603,766 generated bytes in 3.135488 seconds at 48,152,576-byte peak RSS and retained zero output. It has no private cohort-manifest surface. Exact implementation a2e19ba passed Base job 97099725036, Optional job 97099725145, and CI 32601257065. Read the VR24P proof-only closeout; it binds the 7,138-byte preproof registry, its hash and Git blob, the artifact-set hash, and four exact artifact blobs without repeating qualification or touching private state. Commit, push, and green this closeout before any readiness or private operation. Exact closeout 589b0f5 passed Base job 97101585701, Optional job 97101585762, and CI 32602001717 before the sole invocation.

The command then read and strict-parsed exactly 418,755 target-free structural bytes once, called VR23A once with one nested VR20A call, and consumed at MARC2VR24P-R2 without freezing a cohort. R2 identifies only the complete- bundle count arithmetic class against the public 238-bundle assumption. The observed count, difference, predicate, value, row, identity, participant, selection, and cohort were not retained. Do not retry or inspect private state. Neural payload, targets, models, predictions, scores, FW2/CIL1, and scientific claims remain closed.

Historical Frontier: MARC2-VR22P Generated Stage 1

Read docs/MARC_2_PUBLISHED_TASK_PRIVATE_CONFIRMATION_AUTHORIZATION_PACKET.md, registries/marc2_published_task_private_confirmation_authorization_request.v0.json, docs/MARC_2_PUBLISHED_TASK_PRIVATE_CONFIRMATION_REQUEST_PROOF_CLOSEOUT.md, and its machine proof before doing anything else.

The generated-only MARC2-VR20A adapter corrected the inherited task identity from task-freewill to the primary-source-published task-reachingandgrasping. It passed 20 source-exact paths and 53 direct refusals, preserving the generated 16-subject / 96-bundle / 384-member selection. Exact implementation bf4d2b7 passed both jobs in CI 32486620566; proof-only closeout 9b5bea4 passed both jobs in CI 32487854026.

The all-false MARC2-VR20P request bef2391d8edf92c5edf8a3624831e50430636626 passed Base job 96793861959, Optional Neuro Readers job 96793861717, and CI 32489589922. Its immutable proof closeout 88b3b4aaa4436655ce6f4de65215982e2b8ff9de passed Base job 96797011698, Optional Neuro Readers job 96797011783, and CI 32490587975.

The exact packet-bound authorization decision 49b0f4b9b8ec13f885b0010947326fbd3b9641ac passed Base job 96991921763, Optional Neuro Readers job 96991921631, and CI 32556695686 before Stage 1 began.

Generated MARC2VR20P-G1 now passes the frozen six-case matrix in two orders and two replays: 24/24 paths, four copies each of G1 and R2-R6, 24 unchanged VR20A calls, and 91 direct refusals. It selected the expected generated 16-subject / 96-bundle / 384-member cohort on the four accepted paths. Runtime was 1.198090 seconds at 35,504,128-byte peak RSS; peak temporary output was 222,998 bytes and retained output was zero.

Exact Stage 1 implementation 81784e7 passed both jobs in CI 32557573872. Proof-only closeout 7cd0503 then passed Base job 96995623811, Optional Neuro Readers job 96995623848, and CI 32558200426 before the sole registered invocation.

Read docs/MARC_2_PUBLISHED_TASK_PRIVATE_CONFIRMATION_RESULT.md and its machine result. The command collected three readiness samples, opened and strict-parsed exactly 418,755 target-free structural bytes once, called VR20A once, and consumed at MARC2VR20P-R5 without a cohort freeze. Runtime was 0.048262 seconds at 29,573,120-byte peak RSS, output was 2,765 bytes, and network/new-payload bytes were zero.

R5 narrows the remaining blocker to VR20A F06 taxonomy/eligibility arithmetic versus F07 selection/split/rank/reservation arithmetic. It excludes the earlier published-task/identity/path/companion class for this exact run, but does not reveal the failed predicate or any private value. VR20P is consumed; do not retry, reopen, rehash, or inspect its private state. The next lane is frozen as MARC2-VR21A. Read the R5 two-route discriminator preregistration and machine contract. Its 12 generated paths distinguish F06 taxonomy/eligibility from F07 selection/split/rank/reservation arithmetic through unchanged VR20A. Registration 7ce0e16 passed both jobs in CI 32559365362 before implementation. Generated MARC2VR21A-G1 now passes all 12 paths with G1/R1/R2 each appearing four times, 12 unchanged VR20A calls, 48 direct refusals, exact replays, and zero retained output. Read the VR21A implementation document and machine records. Exact implementation 661b18d passed both jobs in CI 32560284291; proof closeout 3aa166a then passed both jobs in CI 32560933310 without repeating qualification or performing a private operation. The all-false MARC2-VR22P packet now freezes one future two-stage F06/F07 discriminator with one possible 418,755-byte structural read. Exact all-false request c90b5ad passed both jobs in CI 32561578590. Its proof-only closeout now binds the unchanged 31,463 request bytes and three Git blobs without performing a private operation. Exact proof closeout 026ecd2 passed both jobs in CI 32562175927. Codex then identified VR22P as the sole active Tier C gate, and the maintainer's next exact message was continue. The packet-bound decision preserves those eight UTF-8 bytes and binds only the unchanged two-stage request. It is ineffective until its exact commit is pushed and both required CI jobs are green. Exact decision 197f253 passed both jobs in CI 32593234295 before Stage 1. Generated MARC2VR22P-G1 now passes 12/12 paths with four copies each of G1/R4/R5, 12 exact VR20A calls, eight conditional VR21A map calls, 121 direct refusals, exact replay, and zero retained payload. Runtime was 0.867054 seconds at 35,946,496-byte peak RSS. Read the implementation document and machine record. The exact implementation bc8ecc3 passed both jobs in CI 32594134914. Read the proof-only closeout; it binds the exact implementation registry and four Git blobs without repeating qualification or touching private state. The closeout itself must now be committed, pushed, and remotely green before the one private structural invocation. FW2, archive payload, neural data, targets, models, scoring, and all five scientific claims remain closed.

The five requested scientific targets and their exact evidence order are in docs/FIVE_CLAIM_PROOF_STRATEGY_2026-08-22.md. Preserve its distinction between closed-set language information and arbitrary thought-to-text.

Historical Frontier: MARC2-VR15P Consumed At Run-Token Width

Read docs/MARC_2_SUFFIX_IDENTITY_PRIVATE_DISCRIMINATOR_RESULT.md and its machine registry. Portable activation a9ebef4 passed Base job 96694803139, Optional Neuro Readers job 96694803152, and CI 32456531938 before the sole registered invocation.

The command collected three fresh readiness samples, opened and strict-parsed exactly 418,755 target-free structural bytes once, called VR15A once with one nested unchanged VR12A call, and consumed at MARC2VR15P-R15. Runtime was 10.096427 seconds at 29,016,064-byte peak RSS, combined output was 2,288 bytes, and network/new-payload bytes were zero.

R15 means only the run-token width class. The real structural source violates VR12A's frozen one-or-two-ASCII-digit run-<index> assumption, but the exact token, filename, path, row, identity, participant, selection, and cohort were not retained. No cohort, archive member, neural payload, signal, target, model, prediction, or score was accessed.

VR15P is consumed and cannot be retried or reinspected. The next safe work is frozen as MARC2-VR16A. Read the variable-width primary-source research, preregistration, and machine contract. BIDS allows arbitrary leading zeroes; the generated adapter therefore accepts [0-9]+ under the existing 1,024-byte name cap while keeping semantic runs 1/2/3 and every companion, collision, identity, split, storage, and privacy guard exact.

The frozen matrix has six width patterns, two source orders, two replays, 24 success paths, and at least 48 direct refusals. Commit, push, and green the registration before implementation. Any private read, cohort freeze, neural access, FW2, or CIL1 requires a new Tier C packet and decision.

Prior Frontier: MARC2-VR13P Consumed Invalid And Parked

Read docs/MARC_2_P15_PRIVATE_CONFIRMATION_RESULT.md, then the VR13A preregistration, implementation, and result documents with their registries.

VR12A's generated-only repair accepts one- or two-digit ASCII run-<index> spellings while preserving source-exact names and all subject/session, task, companion, count, split, rank, and storage gates. Generated route MARC2VR12A-G1 passed 12/12 paths and 36 direct refusals. VR12P then qualified the fixed one-shot wrapper on 12/12 generated paths and 61 refusals. Final implementation d98a011 passed Base job 95903371693, Optional job 95903371721, and CI 32197145780; proof-only closeout 4280aa6 passed Base job 95905146777, Optional job 95905146692, and CI 32197772060 before the sole registered invocation.

The command read and strict-parsed exactly 418,755 target-free structural bytes once, called VR12A once, and consumed at MARC2VR12P-R4. Under the frozen route table, R4 means only that the repaired source still refused in the identity, task, or companion-validation layer. It does not reveal the failed predicate, value, row, path, person, task spelling, companion state, selection, or cohort. No private cohort manifest was created.

VR12P has no retry, rerun, resume, repair, fallback, substitution, or private reinspection. FW2 and CIL1 remain ineligible. No archive member, neural payload, signal sample, target, model, prediction, score, release, or scientific claim is open.

VR13A freezes and generated-qualifies the next safe question without reopening that evidence. It partitions the remaining first-failure surface into seven ordered classes: suffix-bearing BIDS identity, exact task token, mixed run tokens, normalized companion collision, incomplete companions, bundle-total mismatch, and taxonomy/eligibility mismatch. Registration 1177174 passed both jobs in CI 32424688012 before implementation. The generated matrix then exercised one success control plus all seven classes in two orders and two replays for 32 unchanged VR12A calls. Every route appeared four times, both replays matched, 54 direct refusals passed, and zero generated output was retained.

The measured pass processed 13,741,736 generated bytes in 2.401634 seconds at 36,978,688-byte peak RSS and emitted 5,514 aggregate bytes. No private path, archive, neural payload, target, model, score, FW2, or CIL1 operation occurred. Exact implementation 63a0b8e passed Base job 96610793887, Optional job 96610793714, and CI 32426975815. The proof-only closeout repeats no qualification and performs no private operation. Tier A may now prepare one all-false private residual-discriminator packet. Any later private read still requires that exact packet to be remotely green and receive a fresh Tier C packet-bound decision. FW2 and CIL1 remain ineligible.

That packet is now prepared as MARC2-VR13P. It binds 17 committed artifacts totaling 321,145 bytes and requests two proof-separated stages: generated/mock fixed-path wrapper qualification, then one future 418,755-byte target-free structural read. Success R1 may freeze one 12-19-subject structural cohort; R2-R8 retain only one residual blocker class. All authorization fields are false, all operation counters are zero, and no private path was inspected.

Immediate gate: commit, push, and green this request, then make and green a non-scope-changing request proof closeout. Only after VR13P is identified as the sole active Tier C packet can a fresh maintainer message authorize its unchanged two-stage sequence. No earlier continue is retroactive.

Request d55371e passed Base job 96615486644, Optional job 96615486542, and CI 32428583270. Its separate proof closeout binds the three unchanged request artifacts and keeps all authority false. That closeout must now be committed, pushed, and remotely green. Do not implement or inspect readiness, the private source, or an output root before a fresh decision after that proof.

Final proof head bff3d3f passed Base job 96618310916, Optional job 96618311046, and CI 32429569470. After VR13P was identified as the sole active Tier C packet, the maintainer's next exact message was continue. The separate authorization decision quotes those eight bytes and binds only the unchanged packet. Commit, push, and green that decision before implementing the generated wrapper or touching readiness, .codex_work, the private source, or an output root.

Decision fe16400 passed Base job 96648078587, Optional job 96648078452, and CI 32439821302. Stage 1 now implements the fixed plan, qualify, inspect, and proof-gated execute surfaces. Generated route MARC2VR13P-G1 passed all 32 matrix paths, one temporary fixed-path success path, and 81 direct refusals. The measured pass used 14,171,146 generated bytes, 3.208020 seconds, 34,324,480-byte peak RSS, 2,087 aggregate bytes, and zero retained/private/neural/model/score operations.

Initial Stage 1 36556e6 passed both jobs but was superseded before use when pre-closeout audit found its record test accepted only null proof. Final proof-state-hardened implementation 4fb2424 passed Base job 96654430846, Optional job 96654430654, and CI 32442008002.

During local proof-closeout preparation, a focused test expected execute to refuse while readiness and preflight were mocked. The executor trusted the uncommitted local proof fields, returned instead of refusing, and therefore opened and read the 418,755-byte target-free structural source once before the test assertion failed. The separate proof closeout had no commit or remote green proof, so this is an invalid invocation, not evidence.

The returned route was not retained, and the ignored output was not inspected. VR13P is consumed and parked with proof restored to null and no retry, rerun, resume, cleanup, inference, FW2/CIL1 eligibility, or claim. Read the incident document and machine record. Any aggregate-output recovery requires a new frozen Tier C packet and fresh decision.

That all-false recovery request is now frozen as MARC2-VR14P. It binds the green incident head and six tracked artifacts, while authorizing nothing now. The proposed future sequence may open only the existing aggregate report once, at most 65,536 bytes, after a fresh decision plus separate green implementation and proof barriers. The structural source, readiness, marker, private manifest, archive, neural data, targets, models, scores, FW2/CIL1, and claims stay closed.

Exact VR14P packet d920e8eeaf7a7e9c980232c5de59f0e390c374be passed Base job 96657974654, Optional Neuro Readers job 96657974564, and CI 32443248466. Its request proof closeout binds the three unchanged packet artifacts totaling 13,334 bytes and performs zero ignored-path operations.

Immediate gate: commit, push, and green the proof closeout. Only then may VR14P be identified as the sole Tier C gate for new packet-bound maintainer words.

Proof closeout 3274a728ccf25a2e7bb5a7c208d2e8d53f2db6fb passed Base job 96659529617, Optional job 96659529824, and CI 32443804353. After VR14P was identified as the sole gate, the maintainer's next message authorized only that unchanged two-stage packet by reference. The decision explicitly rejects blanket or future authority from its continuous-approval wording.

Immediate gate: commit, push, and green the exact decision. Only then may the generated/mock fixed-path reader be implemented. The aggregate report remains closed until that implementation and a separate proof closeout are green.

Decision 60b97ea6c9715b651c17bb6d797c1f02c10ba9e2 passed both jobs in CI 32444425790 before Stage 1. Generated MARC2VR14P-G1 now passes all 32 route/order/replay paths and 89 direct refusals in 0.008616 seconds at 28,901,376-byte peak RSS. It retained zero output and touched no ignored path.

Immediate gate: commit, push, and green the exact implementation. Then create and green its separate proof closeout. The one real aggregate open remains closed until both barriers pass and explicit one-shot arming is present.

Exact Stage 1 046013a4a8089f5a9f3a91fc246420cac21a1d20 passed both jobs in CI 32445483857. Its proof closeout binds the implementation registry and five owned artifacts totaling 68,625 bytes without repeating qualification or touching an ignored path.

Immediate gate: commit, push, and green the proof closeout. Then create and green the tracked-clean activation proof before the single armed execution.

Proof closeout f352c1b passed both jobs in CI 32446071998. The separate activation record now binds the exact implementation and closeout proofs while performing zero ignored-path operations.

Immediate gate: commit, push, and green the activation record. Then run the single explicitly armed aggregate-only command exactly once.

Activation 6bfff69 passed both jobs in CI 32446635433 before the sole execution. VR14P consumed at MARC2VR13P-R2: the suffix-bearing BIDS identity structural class only. The run read 1,543 aggregate bytes once and wrote a 1,945-byte receipt; source, private-manifest, neural, target, model, and score operations were zero.

Immediate gate: freeze an artifact-only and generated-only suffix-identity decomposition. Do not reopen VR14P or access private/neural material.

MARC2-VR15A now freezes that artifact-only/generated-only decomposition: 15 ordered P15 grammar classes, one multiple-class route, and 68 exact generated adapter paths under a 32 MiB cap.

Immediate gate: commit, push, and green the VR15A registration before any implementation. Private, ignored, neural, target, model, and score work stays closed.

Prior Frontier: MARC2-VR11P Consumed At Structural Route R2

Read docs/MARC_2_MACHINE_STABLE_PRIVATE_RECOVERY_RESULT.md, then docs/MARC_2_VR2_REFUSAL_LOCALIZATION_PREREGISTRATION.md and docs/MARC_2_VR2_REFUSAL_LOCALIZATION_IMPLEMENTATION.md with their registries. The sole VR4P pass is consumed: it verified and strict-parsed exactly 418,755 private structural bytes, called VR2 once, and returned no cohort. Artifact-only VR5A then proved that the wrapper collapsed eight safe nested VR2 routes into one outer MARC2MSP-F07, and that the live selection path still enforces nine exact generated-fixture outputs and generated source semantics. The hidden private route remains unavailable and must not be inferred as F06. Exact implementation d9fa443 passed Base job 95229811925, Optional job 95229811955, and CI 31973656275.

Registration 71d7cec passed both jobs in CI 31974405202 before the generated-only implementation began. Local MARC2VR6-G1 now validates maximal prefixes of 12, 14, 16, 18, and 19 subjects in both row orders, preserves only an allowlisted nested route code, recomputes live selection identities, and passes 34 direct refusal cases. The measured pass used 4,291,124 generated input bytes, 9,954 aggregate output bytes, 1.003502375 seconds, and 40,435,712 bytes peak RSS, with zero retention and every private or scientific counter at zero. Exact implementation 482dad5 passed Base job 95234487830, Optional job 95234487789, and CI 31975600088.

Do not touch any ignored MARC2 path, consumed marker, certificate, output root, or retained source. No new private read, real cohort freeze, archive member, neural payload, target, model, prediction, score, FW2, or CIL1 work is authorized by VR6. That next gate is now specified as MARC2-VR7P in docs/MARC_2_DYNAMIC_PRIVATE_SELECTION_RECOVERY_AUTHORIZATION_PACKET.md and its registry. It binds new readiness/output paths, one future 418,755-byte structural open, one VR6 call, dynamic 12-19-subject outcomes, and 72 generated refusals. Immutable request 9d42bac29b695a97639c4a197812865f0ac4f7d5 passed Base Python job 95236917861, Optional Neuro Readers job 95236917836, and CI 31976595268. It is the sole active Tier C gate, but every authorization field and request-stage counter was false/zero. The maintainer's next exact word was continue, recorded in the separate VR7P decision. Decision a318521 passed Base job 95244335512, Optional job 95244335508, and CI 31979669507 before implementation began.

Read docs/MARC_2_DYNAMIC_PRIVATE_SELECTION_RECOVERY_IMPLEMENTATION.md, docs/MARC_2_DYNAMIC_PRIVATE_SELECTION_RECOVERY_RESULT.md, and their registries. Exact wrapper 154852c passed Base job 95252133987, Optional job 95252133958, and CI 31982672176 before the sole structural invocation. The run crossed readiness, wrote its consumed marker, read and verified exactly 418,755 target-free structural bytes once, strict-parsed once, and called VR6 once. It consumed at MARC2VR7P-F07, preserving only upstream VR6 route MARC2VR6-F02, which identifies VR6's upstream VR2-validation branch. The nested VR2 route, failed private predicate/value, intermediate selection, and real cohort were not retained. Do not reopen or alter the new readiness path, output root, marker, or source, and do not rerun VR7P. FW2, CIL1, neural payload access, training, scoring, and live decoding remain closed. Initial result record ae75423 passed both jobs in CI 31983281390; that was test and record proof only, not a second private execution.

Read docs/MARC_2_VR6_VR2_BOUNDARY_LOCALIZATION_PREREGISTRATION.md, docs/MARC_2_VR6_VR2_BOUNDARY_LOCALIZATION_IMPLEMENTATION.md, and their registries. Registration d33eaf3 passed Base job 95256950555, Optional job 95256950656, and CI 31984475999 before implementation. The artifact-only audit reached MARC2VR8A-R1: exact producer lineage excludes VR2 envelope route F02, leaving only F03 path/run-companion structure and F04 bundle/ taxonomy arithmetic compatible with the consumed outer branch. VR6 stored the nested allowlisted route in upstream_route, while VR7P forwarded only the outer route. Generated VR2 success starts from a selector-authored 1,227-row manifest and does not traverse the exact producer, whose parser fixture has only 18 rows. The exact private F03/F04 route remains unavailable. Exact implementation/result 1addd5d passed Base job 95261271737, Optional job 95261271709, and CI 31986089529.

Read docs/MARC_2_GENERATED_DIAGNOSTIC_RELAY_PREREGISTRATION.md, docs/MARC_2_GENERATED_DIAGNOSTIC_RELAY_IMPLEMENTATION.md, docs/MARC_2_GENERATED_DIAGNOSTIC_RELAY_RESULT.md, and their registries. Registration 5607fe8 passed Base job 95263869003, Optional job 95263869149, and CI 31987093865 before implementation. Generated route MARC2VR8B-G1 now sends all 1,227 entries through the exact parser and live manifest producer before VR2/VR6. Canonical and reversed success plus forced F02/F03/F04 cases preserve only outer MARC2VR6-F02 and the exact nested VR2 code. Two full replays pass all 16 gates and 29 direct refusals over 4,650,480 generated bytes in 2.421215 seconds at 59,310,080-byte peak RSS with zero retention and zero forbidden operations.

The generated 16-subject/96-bundle selection is identical across source orders, while VR6's existing provenance hash correctly differs because it binds ordered source rows. VR8B therefore records a separate order-neutral cohort digest without changing VR6. After 3,948 dependency-light and 4,019 optional tests pass with zero new failures, exact implementation d7ce48b passed Base job 95271230358, Optional job 95271230485, and CI 31989817593. Tier A/B may prepare the next all-false packet, but do not infer whether the consumed private source reached F03 or F04 and do not relax either class. Any new private structural read or real cohort freeze remains a fresh Tier C packet and decision; FW2, CIL1, neural payload access, training, scoring, and live decoding remain closed.

That all-false packet is now prepared in docs/MARC_2_TWO_LAYER_PRIVATE_DIAGNOSTIC_AUTHORIZATION_PACKET.md and registries/marc2_two_layer_private_diagnostic_authorization_request.v0.json. VR9P binds 17 green predecessor artifacts totaling 328,581 bytes and proposes one future two-stage diagnostic. Request de8e6dc passed Base job 95277554517, Optional job 95277554619, and CI 31992178980; proof closeout ddc5e85 passed Base job 95278576871, Optional job 95278576905, and CI 31992563746. After Codex identified VR9P as the sole active Tier C packet, the maintainer supplied the exact message continue. Read docs/MARC_2_TWO_LAYER_PRIVATE_DIAGNOSTIC_AUTHORIZATION_DECISION.md and its registry. Decision 4cdd3d3 passed Base job 95280728093, Optional job 95280728134, and CI 31993388608 before implementation began. Now read docs/MARC_2_TWO_LAYER_PRIVATE_DIAGNOSTIC_IMPLEMENTATION.md, its registry, proof record, and tests. The generated/mock fixed-path wrapper passes F03/F04 in both orders across two exact replays, eight VR6 calls, and 70 direct refusals. Exact implementation 0dd113a passed Base job 95285174846, Optional job 95285174911, and CI 31995078475 before the sole execution.

Read docs/MARC_2_TWO_LAYER_PRIVATE_DIAGNOSTIC_RESULT.md and its registry. The one run crossed three passing readiness samples, read and strict-parsed exactly 418,755 target-free structural bytes once, and called VR6 once. It retained aggregate route MARC2VR9P-R1, outer MARC2VR6-F02, and nested MARC2VR2-F03. This distinguishes F03 path/companion/grouping structure from F04 taxonomy/arithmetic, but reveals no failed predicate, value, row, path, identity, candidate, selection, or cohort. Runtime was 10.044833041 seconds at 39,075,840-byte peak RSS; all neural, target, model, training, prediction, score, network, and other-project counters were zero.

VR9P is consumed with no retry or private reinspection. FW2 and CIL1 remain closed because no cohort exists. Next perform only a separately named artifact-only F03 predicate-decomposition and generated-coverage pass. Any further private read needs a new immutable packet and fresh Tier C decision.

That pass is now preregistered as MARC2-VR10A. Read docs/MARC_2_F03_PREDICATE_DECOMPOSITION_PREREGISTRATION.md and its contract. The exact F03 class contains 20 leaf predicates. Fifteen are excluded by committed producer invariants or the live aggregate counts; five remain unresolved: maximum path length, suffix-bearing BIDS identity, exact task token, logical companion uniqueness, and four-companion completeness. The contract freezes a six-case full-scale generated witness matrix, canonical and reversed order, two exact replays, 24 VR6 calls, 29,448 parser-entry visits, and at least 40 refusals. It is artifact-only and generated-only. No unresolved predicate is asserted to be the private cause, and no private, archive, neural, target, model, score, FW2/CIL1, or scientific operation is authorized.

Registration 80175a7 passed both jobs in CI 31997129703 before the generated implementation began. Now read docs/MARC_2_F03_PREDICATE_DECOMPOSITION_IMPLEMENTATION.md and docs/MARC_2_F03_PREDICATE_DECOMPOSITION_RESULT.md. Local route MARC2VR10A-G1 verifies all 20 AST signatures and sends one full-scale witness for each unresolved class through the exact parser, producer, VR2, and VR6. Across two orders and two replays, four controls succeed and all 20 witness paths return outer F02 plus nested F03. The pass used 6,979,708 generated bytes in 1.8363693330029491 seconds at 45,072,384-byte peak RSS, retained zero, and passed 47 direct refusals. All 31 focused and 4,042 complete local tests pass. Exact implementation 84103a5 passed Base job 95295212461, Optional job 95295212440, and CI 31998811585. This result narrows the engineering hypothesis set but does not identify the private cause or establish neural or decoding evidence. The next safe work is a separately frozen generated-only five-route discriminator; a private invocation remains closed.

That discriminator is now implemented locally as MARC2-VR10B. Read docs/MARC_2_F03_FIVE_ROUTE_DISCRIMINATOR_PREREGISTRATION.md and registries/marc2_f03_five_route_discriminator_contract.v0.json. Registration d642eae passed Base job 95308775711, Optional job 95308775577, and CI 32003674374 before implementation. Then read docs/MARC_2_F03_FIVE_ROUTE_DISCRIMINATOR_IMPLEMENTATION.md, docs/MARC_2_F03_FIVE_ROUTE_DISCRIMINATOR_RESULT.md, and their registries. The ordered first-match classifier maps P03, P15, P16, P18, and P19 to R1-R5; G1 is generated-control only. Across six exact-parser cases, two orders, and two replays, every route appeared exactly four times. The pass completed 24 parser/producer paths, 24 VR6 calls, 24 discriminator calls, and 29,448 parser- entry visits; it passed 60 direct refusals, used 6,979,708 generated bytes in 2.725759166991338 seconds at 44,564,480-byte peak RSS, emitted 7,515 aggregate bytes, and retained zero. Thirty-two focused and all 4,074 base tests pass with 204 expected skips. Exact implementation/result 61bb801 passed Base job 95320325187, Optional job 95320325136, and CI 32007641751; no qualification or private operation was repeated for proof closeout. No private executor, private read, FW2/CIL1 action, neural payload, target, model, prediction, score, or scientific claim is authorized. A future private discriminator remains a new Tier C packet and fresh decision.

That all-false packet is now prepared as MARC2-VR11P. Read docs/MARC_2_F03_PRIVATE_DISCRIMINATOR_AUTHORIZATION_PACKET.md, registries/marc2_f03_private_discriminator_authorization_request.v0.json, and its request test. It binds 16 tracked predecessors totaling 295,028 bytes and proposes one future generated/mock wrapper followed, only after separate green decision and implementation barriers, by one 418,755-byte target-free structural read, one VR6 call, one VR10B call, and one aggregate R1-R5 route. Every authority flag is false and every operation counter is zero. The request commit 6e72c8f797201359777454a750b1dea9704665c0 passed Base Python job 95326004060, Optional Neuro Readers job 95326004145, and CI 32009557248. Proof closeout 136f7b9 then passed Base Python job 95330380822, Optional Neuro Readers job 95330380918, and CI 32011020786. After Codex identified VR11P as the sole active Tier C packet, the maintainer's next message was exactly okay lets continue. Read docs/MARC_2_F03_PRIVATE_DISCRIMINATOR_AUTHORIZATION_DECISION.md, its registry, and its test. The decision quotes those 18 UTF-8 bytes and binds only the unchanged two-stage packet. Decision 4fa2771 passed Base job 95414004791, Optional job 95414004814, and CI 32038683203 before Stage 1 began.

Now read docs/MARC_2_F03_PRIVATE_DISCRIMINATOR_IMPLEMENTATION.md, its implementation registry, and its test. The generated/mock wrapper passes 24 exact VR6/VR10B paths, one temporary exact-size fixed-path state-machine pass, and 81 direct refusals. Each G1/R1-R5 route appears four times. The measured qualification used 7,398,463 generated bytes, 3.424092 seconds, 63,619,072-byte peak RSS, 7,268 aggregate bytes, and zero retained output or forbidden operations. The clean base suite passes 4,117 tests with 204 expected skips.

After proof-state hardening, exact implementation 2093ad5 passed Base job 95421634020, Optional job 95421633971, and CI 32041540553. The registry now binds that exact proof without rerunning qualification or touching a private path. This proof-only closeout must itself be committed, pushed, and green in both jobs before execute may be used.

Proof-only closeout e569bcc passed Base job 95422480212, Optional job 95422480363, and CI 32041863346 before the sole Stage 2 invocation. Read docs/MARC_2_F03_PRIVATE_DISCRIMINATOR_RESULT.md and its result registry. The run crossed readiness, read exactly 418,755 target-free structural bytes once, called VR6 once and VR10B once, and consumed at MARC2VR11P-R2. Under the frozen route table, R2 means only that the manifest reaches P15, the suffix- bearing BIDS identity structural class. Runtime was 10.041579 seconds at 34,701,312-byte peak RSS; no archive member, neural payload, target, model, prediction, or score was accessed.

VR11P has no retry, rerun, resume, repair, fallback, substitution, or private reinspection. The next safe work is a separately frozen artifact-only and generated-only P15 repair design. FW2, CIL1, neural-data access, training, scoring, live decoding, and scientific claims remain closed.

Read README.md for the mission, results, proof boundaries, and quickstart.
Read CONTRIBUTING.md for the EEG data/hardware contribution paths.
Read AGENTS.md for coding-agent rules.
Read docs/MARC_2_CONFOUND_TRIANGULATION_RESEARCH.md, its registry, docs/MARC_2_FREEWILL_PREFIX_SELECTION_PREREGISTRATION.md, and its registry for the current five-work-order scientific path and first frozen contract. Then read docs/MARC_2_FREEWILL_PREFIX_SELECTION_IMPLEMENTATION.md, docs/MARC_2_FREEWILL_PREFIX_SELECTION_SYNTHETIC_RESULT.md, and their registries. The prior Wrist branch is consumed and blocked, and the one generated MARC2-FW1 closeout is also consumed. Then read docs/MARC_2_FREEWILL_PRIVATE_SELECTION_AUTHORIZATION_PACKET.md and its registry, then read docs/MARC_2_FREEWILL_PRIVATE_SELECTION_AUTHORIZATION_DECISION.md and its registry. Decision ad1e406 is remotely green. Then read docs/MARC_2_FREEWILL_PRIVATE_SELECTION_IMPLEMENTATION.md and registries/marc2_freewill_private_selection_implementation.v0.json. The generated/mock wrapper passes all 58 refusal checks and the complete local suites with zero retained-private operations. Exact implementation d9a3853 passed both remote jobs before its sole execution. Then read docs/MARC_2_FREEWILL_PRIVATE_SELECTION_RESULT.md and registries/marc2_freewill_private_selection_failure_result.v0.json. Execution failed closed at MARC2FWS-F00 because the committed implementation registry omitted its required top-level lane_id. The failure occurred before any retained path check, marker, source open, read, hash, parse, selection, or output. MARC2-FW1A is consumed without rerun. Then read docs/MARC_2_PROOF_RECORD_RECOVERY_PREREGISTRATION.md and registries/marc2_proof_record_recovery_contract.v0.json. The separately named MARC2-FW1B lane freezes a generated-only shared validator, 15 exact implementation-record fields, and 32 refusal cases. It has no execute command and a private execution limit of zero. The contract must be pushed and pass both remote jobs before its generated implementation begins. Exact contract b86aa94 passed both jobs in CI 31767373647. Then read docs/MARC_2_PROOF_RECORD_RECOVERY_IMPLEMENTATION.md, its registry, module, and two implementation test files. Generated MARC2FWR-G1 passed two exact canonical validations and all 32 refusal cases through one shared validator in 0.016927 seconds at 27,099,136-byte peak RSS. Private execution limit and every private, real, network, neural, target, model, and score counter remain zero. Exact implementation 6f613b3 passed both remote jobs in CI 31768593977. Then read docs/MARC_2_LIVE_SELECTION_RECOVERY_AUTHORIZATION_PACKET.md, registries/marc2_live_selection_recovery_authorization_request.v0.json, and its request test. MARC2-FW1C proposes a generated/mock wrapper followed only after separate green proof by one target-free structural-manifest selection. Exact request 7804c3e passed both jobs in CI 31769518851 and was identified as the sole Tier C gate. Then read docs/MARC_2_LIVE_SELECTION_RECOVERY_AUTHORIZATION_DECISION.md, registries/marc2_live_selection_recovery_authorization_decision.v0.json, and its decision test. The decision quotes the maintainer's fresh exact continue and binds only the packet. Exact decision b0466e5 passed both jobs in CI 31928627432, opening generated/mock Stage 1 only. Then read docs/MARC_2_LIVE_SELECTION_RECOVERY_IMPLEMENTATION.md, registries/marc2_live_selection_recovery_implementation.v0.json, registries/marc2_live_selection_recovery_proof_certificate.v0.json, the additive module, and both recovery tests. The native FW1C registry and distinct FW1B certificate pass the exact shared validator; the generated closeout passed all 90 proof/selector/wrapper mutations in 0.374161 seconds at 38,666,240-byte peak RSS with zero private operations. Exact wrapper 7b924be passed both jobs in CI 31930051249 before its sole Stage 2 invocation. Then read docs/MARC_2_LIVE_SELECTION_RECOVERY_RESULT.md, registries/marc2_live_selection_recovery_failure_result.v0.json, and the result test. One exact 418,755-byte structural read passed size, hash, and strict JSON, then consumed at MARC2FWC-F02 because strict live source identity differed. Zero participants or members were selected, no private selection was written, no retry is open, and MARC2-FW2 is ineligible. Then read docs/MARC_2_SOURCE_SCHEMA_LINEAGE_AUDIT.md and its result registry. Artifact-only MARC2-SL1 found the exact integration defect: the producer emits transport key directory, while the generated fixture and both consumer validators require central_directory. This explains F02 without private reinspection. Then read the MARC2-TA1 preregistration, implementation document, and two implementation/result registries. Green registration commit 0c0e1c8 preceded the generated-only adapter. One measured qualification passed all 26 mutations and replayed the unchanged 16-subject/384-member generated selector identity in 0.453316 seconds at 39,108,608-byte peak RSS. Exact implementation/result commit 108b869 passed Base job 95132260089, Optional job 95132260076, and CI 31933692066. No temporary output, live/private read, neural payload, target, model, score, or network operation remains. Archive members, payloads, signals, targets, models, scores, language datasets, provider calls, and MARC2-FW2 remain unauthorized. Then read docs/MARC_2_LIVE_SCHEMA_ADAPTER_PREREGISTRATION.md and its contract. MARC2-LA1 freezes the next generated/mock composition: validate the exact live-shaped envelope, bridge only four identity values into TA1, call its green public adapter once, and replay the frozen selector under 30 mutations. Registration 62e465e passed CI 31934737967 before the generated implementation. Then read docs/MARC_2_LIVE_SCHEMA_ADAPTER_IMPLEMENTATION.md and both result registries. One MARC2LA-G1 run passed all mutations in 0.488921 seconds at 38,387,712-byte peak RSS; its 5,366-byte report was removed and all private/neural/model counters stayed zero. Exact implementation 3e3f8b8 passed Base job 95137289730, Optional job 95137289704, and CI 31935754822; proof closeout 52eabc6 passed CI 31936399968. Then read docs/MARC_2_LIVE_ADAPTER_RECOVERY_AUTHORIZATION_PACKET.md, its request, and the separate authorization decision. Request f9f24a3 passed Base job 95140483613, Optional job 95140483638, and CI 31937038394 before the maintainer supplied the fresh exact word continue. The decision binds only that green packet. Decision b445df2 then passed Base job 95142233426, Optional job 95142233335, and CI 31937743296 before the additive executor was implemented. Then read docs/MARC_2_LIVE_ADAPTER_RECOVERY_IMPLEMENTATION.md, its native implementation registry, distinct FW1B proof certificate, module, and three tests. Generated MARC2LAR-G1 passed all 32 proof-certificate and 24 executor refusals, replaying the 16-subject, 96-run, 384-member structural prefix in 0.273685 seconds at 37,978,112-byte peak RSS. Every retained-path, archive, neural, target, model, score, network, and claim counter stayed zero, and the generated output was removed. Exact implementation 5390e06 then passed Base job 95146470514, Optional job 95146470539, and CI 31939483560 before its sole structural execution. Read docs/MARC_2_LIVE_ADAPTER_RECOVERY_RESULT.md and registries/marc2_live_adapter_recovery_failure_result.v0.json. One no-follow pass opened, read, hashed, and strict-parsed exactly 418,755 bytes, then consumed at MARC2LAR-F02 because LA1 refused the source. LA1 recorded no success, the selector was never called, and zero subjects or members were selected. The exact nested LA1 predicate is unavailable without forbidden reinspection. No archive, neural, target, model, score, or network operation occurred. LA2 has no retry; MARC2-FW2 remains ineligible. Then read docs/MARC_2_VALIDATION_COVERAGE_LOCALIZATION.md and registries/marc2_validation_coverage_localization_result.v0.json. The fixed artifact-only audit found that the generated fixture represented only 195 eligible runs while the published source space contains 238, leaving 43 run slots and 172 companion slots untested as Freewill-shaped rows. LA1 applies its exact 195-run equality before eligibility filtering. This is a concrete generated-coverage defect consistent with the F02 class, not the exact private predicate and not neural evidence. Then read docs/MARC_2_SOURCE_VALIDITY_ELIGIBILITY_REPAIR_PREREGISTRATION.md and registries/marc2_source_validity_eligibility_repair_contract.v0.json. MARC2-VR1 freezes a generated 238-bundle source domain containing the 195 eligible bundles plus all 43 omitted slots as source-shaped adversaries. It requires source validation, eligibility classification, filtering, and only then the exact 195-bundle assertion. Registration 9dedfe6 passed CI 31942316544; the standard-library implementation now passes its 36 focused registration/behavior/record tests and preserves the frozen selection with zero ineligible candidates. Read docs/MARC_2_SOURCE_VALIDITY_ELIGIBILITY_REPAIR_IMPLEMENTATION.md and its implementation registry next. Exact implementation 4d587df then passed CI 31943437003 before one measured generated qualification passed at MARC2VR-G1. Read docs/MARC_2_SOURCE_VALIDITY_ELIGIBILITY_REPAIR_RESULT.md and its result registry. Result commit 05fc2b5 passed both remote jobs in CI 31943963317. The generated closeout is consumed with no rerun. Every private or scientific operation remains closed behind a new prospective contract and fresh Tier C decision. Then read docs/MARC_2_LIVE_DOMAIN_ELIGIBILITY_ADAPTER_PREREGISTRATION.md and registries/marc2_live_domain_eligibility_adapter_contract.v0.json. MARC2-VR2 prospectively removes VR1's constructed 12/24/7 ineligible breakdown from live acceptance. It freezes only the exact 238 total bundles, 195 eligible bundles, 43 ineligible bundles, public participant taxonomy, and unchanged target-free selection. Four generated breakdowns in two row orders must all produce the same selection identity. The registration is generated-only and implementation must wait for its own remotely green commit. Registration 384373e passed Base job 95159734989, Optional job 95159734967, and CI 31945086852 before the generated-only implementation began. Read docs/MARC_2_LIVE_DOMAIN_ELIGIBILITY_ADAPTER_IMPLEMENTATION.md and its implementation registry. All eight profile/order paths preserve the same target-free selection and all 58 mutations refuse. Exact implementation f62a3f5 passed Base job 95162220059, Optional job 95162220159, and CI 31946112252 before one registered generated closeout. Read docs/MARC_2_LIVE_DOMAIN_ELIGIBILITY_ADAPTER_RESULT.md and its result registry. MARC2VR2-G1 passed all eight profile/order paths and 58 refusals in 0.5122641660127556 seconds at 32,620,544-byte peak RSS, with zero retained output and every private, neural, target, model, network, and score counter at zero. The generated closeout is consumed with no rerun. Result commit 7b6899b passed both required jobs in CI 31946852669, completing remote proof. No private path, archive payload, neural data, target, model, score, network operation, or FW2 work is authorized. Then read docs/MARC_2_VARIABLE_DOMAIN_PRIVATE_RECOVERY_AUTHORIZATION_PACKET.md and registries/marc2_variable_domain_private_recovery_authorization_request.v0.json. Request commit 328faa8 passed Base Python job 95166799271, Optional Neuro Readers job 95166799305, and CI 31947928896. Then read docs/MARC_2_VARIABLE_DOMAIN_PRIVATE_RECOVERY_AUTHORIZATION_DECISION.md and its decision registry. The maintainer's fresh three-line instruction is preserved verbatim and binds only the green MARC2-VR3 packet. The decision commit 944b6e8 passed Base job 95202667384, Optional job 95202667483, and CI 31962561043 before implementation. Then read docs/MARC_2_VARIABLE_DOMAIN_PRIVATE_RECOVERY_IMPLEMENTATION.md, the native implementation registry, and the distinct FW1B proof certificate. The generated wrapper passed eight profile/order paths and 64 direct mutations in 1.850114875 seconds at 38,322,176-byte peak RSS and emitted 4,464 bytes; every private, payload, neural, target, model, score, and claim counter was zero. Exact implementation 2467876 passed Base job 95207398015, Optional job 95207398092, and CI 31964473405 before its sole invocation. Then read docs/MARC_2_VARIABLE_DOMAIN_PRIVATE_RECOVERY_RESULT.md and its failure registry. The invocation consumed at MARC2VDR-F01 during aggregate machine preflight, before any output-root or private-path operation. Private input, output, cohort, payload, model, and score counters are zero. The exact failed load/disk/RSS predicate was not emitted and must not be inferred. VR3 has no retry or rerun, no real cohort exists, and FW2 remains ineligible pending a separately named structural recovery and fresh Tier C decision. Result commit a186486 passed Base job 95208692240, Optional job 95208692213, and CI 31964995980. Then read docs/MARC_2_MACHINE_STABLE_COHORT_AND_NEURAL_CONTROL_RESEARCH.md and its registry. The prospective order is VR4 machine-stable cohort freeze, FW2 bounded member/semantic qualification, then CIL1 target-firewalled neural controls. Signal, target-independent derangement, timing, and no-signal are explicit, but candidate effect thresholds are not preregistered and every private, payload, training, target, scoring, and claim authority remains false. Then read docs/MARC_2_MACHINE_STABLE_STRUCTURAL_RECOVERY_PREREGISTRATION.md and its contract. VR4 freezes a generated-only readiness certificate with three passing samples, exact diagnostics, 600-second wait, 300-second expiry, 36 mutations, and no execute/private-path interface. Registration 3af2e3d passed Base job 95210732393, Optional job 95210732329, and CI 31965823863 before implementation. Then read docs/MARC_2_MACHINE_STABLE_STRUCTURAL_RECOVERY_IMPLEMENTATION.md, its implementation registry, module, and two focused tests. The generated qualification passed three success paths, one non-ready timeout shape, deterministic replay, and all 36 mutations in 0.007058667 seconds at 19,038,208-byte peak RSS, retaining zero output and performing zero private or scientific operations. The exact implementation must be committed, pushed, and both remote jobs green before one machine-only readiness closeout. The future private open remains a separate Tier C gate; a real cohort, FW2, CIL1, training, targets, scoring, and claims remain closed. Exact implementation 9fdda31 then passed Base job 95213934048, Optional job 95213934126, and CI 31967145837. Read docs/MARC_2_MACHINE_READINESS_RESULT.md and its result registry. One machine-only closeout passed three samples and wrote a 4,551-byte mode-0600 certificate while every private and scientific counter remained zero. The certificate is transient, no real cohort identity exists, and FW2/CIL1 are still ineligible. The result must become remotely green before one separate all-false Tier C structural packet is prepared. Result 0a4a7fb passed Base job 95214802865, Optional job 95214802846, and CI 31967501519. Then read docs/MARC_2_MACHINE_STABLE_PRIVATE_RECOVERY_AUTHORIZATION_PACKET.md and its request registry. MARC2-VR4P is all false and requests only one future generated/mock implementation plus one later proof-gated structural read. It must be committed, pushed, and remotely green before it can be identified as the sole Tier C packet; no current or earlier continue is retroactive authority. Request a5b73d6 then passed Base job 95215825208, Optional job 95215825263, and CI 31967933217. Read docs/MARC_2_MACHINE_STABLE_PRIVATE_RECOVERY_AUTHORIZATION_DECISION.md and its decision registry. The maintainer's fresh continue is recorded verbatim and binds only that green packet. Decision eac3726 passed Base job 95218521665, Optional job 95218521647, and CI 31969063955. Then read docs/MARC_2_MACHINE_STABLE_PRIVATE_RECOVERY_IMPLEMENTATION.md and its implementation/proof registries. The fixed-path executor has passed two generated replays and 27 refusal cases without touching the expired certificate or retained manifest. Commit, push, and green that exact implementation before its sole real structural pass. Payload, neural, training, target, prediction, scoring, FW2/CIL1 execution, and claim surfaces remain closed. Implementation 24f7379 then passed Base job 95223010696, Optional job 95223010631, and CI 31970865212. Read docs/MARC_2_MACHINE_STABLE_PRIVATE_RECOVERY_RESULT.md and its failure registry. The one pass consumed at MARC2MSP-F07 after fresh readiness, marker creation, one exact 418,755-byte structural open, and strict JSON integrity; the frozen VR2 adapter refused before a cohort was returned. Do not reopen or alter the fresh certificate, marker, output root, or retained source. There is no retry, no real cohort, and no FW2/CIL1 eligibility. Failure result 6fa6865 passed Base job 95224622453, Optional job 95224622466, and CI 31971526419. Continue only with artifact-only mismatch localization; a new private read remains Tier C. Then read docs/CODEX_HANDOFF.md for the exact handoff state.
Read docs/LOOP_24_LOCAL_PRECISION_RUNTIME.md for the measured parked result, then use its preregistration and authorization documents for audit.
Read docs/LOOP_25_CAUSAL_PREPROCESSING_RESULT.md and registries/loop25_causal_preprocessing_result.v1.json for the completed target-free mechanics result; use the amendment and authorization records only to audit its frozen execution order. No rerun is authorized.
Read docs/LOOP_26_SHARED_VALIDATION_RESULT.md and registries/loop26_shared_validation_result.v0.json first for the consumed negative result: candidate macro CER 0.938177 versus prior 0.751235, with Loops 26/31/33 parked and no rerun authorized. Then read docs/LOOP_26_SHARED_VALIDATION_PREREGISTRATION.md, docs/LOOP_26_AUTHORIZATION_DECISION.md, and docs/LOOP_26_SHARED_VALIDATION_IMPLEMENTATION.md to audit the green shared design, one-time authorization, implementation, and target-firewall order. Then read docs/LOOP_48_FAILURE_LOCALIZATION_RESULT.md and registries/loop48_failure_localization_result.v0.json for the consumed artifact-only result. One Stage A execution verified four committed JSON artifacts totaling 155,545 bytes and selected descriptive F5 output-distribution instability in 0.016568875 seconds at 23,429,120-byte peak RSS. This is not a proven root cause and no rerun is authorized. Read docs/LOOP_48_PRIMARY_SOURCE_RESEARCH.md, the contract, request, and authorization decision only to audit the frozen pre-execution sequence. Read docs/LOOP_48_TRAIN_ONLY_HYPOTHESIS_PORTFOLIO.md as the immutable five-hypothesis starting point, then read docs/LOOP_48_HYPOTHESIS_DISCRIMINATION_RESEARCH.md and registries/loop48_hypothesis_discrimination.v0.json for the additive six-hypothesis refinement. Then read docs/LOOP_48_TRAIN_ONLY_DISCRIMINATION_PREREGISTRATION.md and registries/loop48_train_only_discrimination_contract.v0.json for the exact Stage B 44/11 split, 20-fit, 35-inference, 41-prediction protocol. Then read docs/LOOP_48_STAGE_B_RESULT.md and registries/loop48_train_only_discrimination_result.v0.json for the consumed result. The primary candidate reached macro CER 0.953566 versus prior 0.822045; all six full-size causal/linear probes were finite and stable but none cleared the prior rule. H4 stable nonseparability is supported, registered fixed-shift H3 has evidence against it, and H1/H2/H5/H6 remain unresolved. Freeze commit 00215b1 passed push CI 29461934145 and PR CI 29461935560 before the 11 train-check targets opened once. Route L50-R05 parks S24 acquisition for this model family. All 55 rows were used historically, so this is E2 diagnostic evidence, not independent validation or neural advantage; no rerun is authorized. Use docs/LOOP_48_STAGE_B_AUTHORIZATION_DECISION.md, the authorization packet, request, implementation, and hash-only freeze only to audit the pre-execution order. Finally, read docs/LOOP_48_STAGE_C_REPRESENTATION_REPAIR_RESEARCH.md and registries/loop48_stage_c_representation_repair_research.v0.json. They select R1 temporal-context starvation as the next falsifiable hypothesis, freeze a 7,692-parameter 470 ms causal candidate against a 7,568-parameter zero-context ablation, and permit only a bounded synthetic mechanics milestone next. No protected cache, real model run, S24/S25 operation, or scientific claim is authorized by that research. Then read docs/LOOP_48_STAGE_C_SYNTHETIC_IMPLEMENTATION.md and registries/loop48_stage_c_synthetic_implementation.v0.json for the exact models, fixture, gate, CLI, zero-update qualification, and source hashes. Then read docs/LOOP_48_STAGE_C_SYNTHETIC_RESULT.md and registries/loop48_stage_c_synthetic_result.v0.json. After correction commit 2836ecc became remotely green, the one synthetic gate ran and is consumed: candidate CER 0.433333, 1/8 exact, versus ablation CER 1.000000. The 0.566667 relative improvement passed, but both absolute gates failed. Stage C is parked without rerun; no real signal, target, cache, S24, or S25 path opened.
Read docs/LOOP_27_PRIMARY_SOURCE_RESEARCH.md for the selected S25 metadata candidate and every blocker that still prevents preregistration, download, signal/target access, or transfer evaluation.
Read docs/LOOP_28_PRIMARY_SOURCE_RESEARCH.md for the strict zero-shot transfer taxonomy, 48-row/0.05-CER/65,535-assignment final-only rule, and the reason calibrated transfer needs a different physical design. It is planning research only; Loop 28 remains Not Started.
Read docs/LOOP_29_PRIMARY_SOURCE_RESEARCH.md for the local-first scalp EEG lane, partner/lab OPM-MEG lane, 15-field modality matrix, six-level qualification ladder, and exact 5-10 GB capacity boundary. It is planning research only; Loop 29 remains Not Started.
Read docs/LOOP_53_PRIMARY_SOURCE_RESEARCH.md, docs/LOOP_53_FRESH_EEG_ACQUISITION_PREREGISTRATION.md, and registries/loop53_fresh_eeg_acquisition_contract.v0.json for the exact four-file, 96,090,264-byte S20 acquisition-only gate. Then read docs/LOOP_53_AUTHORIZATION_DECISION.md, docs/LOOP_53_ACQUISITION_IMPLEMENTATION.md, and docs/LOOP_53_ACQUISITION_RESULT.md with their matching registries. The one invocation passed in 3.629499 seconds at 63,225,856-byte peak RSS, all four source identities matched, and Loop 53 is consumed with no rerun. Header, marker, signal, MAT, target, cache, split, model, training, inference, and scoring counters remain zero. Then read docs/LOOP_54_PRIMARY_SOURCE_RESEARCH.md and registries/loop54_eeg_trial_geometry_research.v0.json for the prospective four-stage VHDR, target-blind signal-quality, isolated VMRK+MAT, and aggregate-closeout firewall. Planning commit aec440a passed the complete local 996-test suite with three expected skips. Loop 53 acquisition is now complete, but Loop 54 is experimentally Not Started: no S20 content stage, split, model, training, inference, score, or scientific claim is authorized. L54-A is now preregistered at commit c114623 in docs/LOOP_54_STAGE_A_VHDR_PREREGISTRATION.md and registries/loop54_stage_a_vhdr_contract.v0.json. It binds one exact 11,705-byte VHDR, no sibling resolution, 18 gates, 22 refusals, one content open, and one registered execution under a 30-second/256 MiB/1 MiB envelope. Then read docs/LOOP_54_STAGE_A_REGISTRATION_CI_RECOVERY.md and registries/loop54_stage_a_registration_ci_recovery.v0.json. Exact-commit run 31127199848 was retried: the neuro job passed, while Base Python installed Ruff 0.16.2 from the old floating requirement and failed on 400 later lint findings. The frozen payload is byte-identical at pinned commit 2232993, whose CI passed both jobs, and exact-tree replay under Ruff 0.15.20 passed 1,095 tests with three skips. Then read docs/LOOP_54_STAGE_A_RECOVERY_AUTHORIZATION_PACKET.md and registries/loop54_stage_a_recovery_authorization_request.v1.json. This is the recovery-bound exact decision surface. Every authorization flag remains false in those immutable snapshots. Request commit 19813a8 passed CI 31283297030; then read docs/LOOP_54_STAGE_A_RECOVERY_AUTHORIZATION_DECISION.md and registries/loop54_stage_a_recovery_authorization_decision.v1.json. The exact user decision is recorded at commit 2177b36, whose CI 31286428489 passed both jobs before implementation. Then read docs/LOOP_54_STAGE_A_VHDR_IMPLEMENTATION.md and registries/loop54_stage_a_vhdr_implementation.v0.json. The strict standard-library parser, dry-run-first CLI, metadata inspector, no-follow source verifier, exclusive bounded outputs, and all 22 refusal classes are synthetic-qualified. Exact implementation b486fdf passed CI 31287819503 before the one registered execution. Then read docs/LOOP_54_STAGE_A_VHDR_RESULT.md and registries/loop54_stage_a_vhdr_result.v0.json. The exact 11,705-byte VHDR passed no-follow, size, Git-blob, and strict-decode checks, then parked at L54A-F11 because the frozen format preamble requirement failed. No raw first line or header field was published, no sibling was touched, and no registered output was written. L54-Q2 failed, the execution is consumed with no rerun, and Loop 54-B/C remain blocked. Then read docs/LOCAL_EEG_TOOLING_AUDIT_2026-08-08.md, registries/local_eeg_tooling_audit_result.v0.json, registries/local_eeg_tooling_audit_receipt.v0.json, and docs/NEXT_20_SYSTEMATIC_EXECUTION_2026-08-08.md. Green implementation e1de855 added a fixed zero-network optional-tool audit; CI 31277731869 passed before the measured run. NumPy/SciPy plus MNE BrainVision reading and ICA are ready locally, while scikit-learn, pyRiemann, MOABB, and Braindecode are absent. Runtime was 14.52799025 seconds, maximum child RSS was 173,211,648 bytes, retained output was 9,416 bytes, and every data, target, model, training, scoring, network, provider, and hardware counter stayed zero. Work orders 1-2 are complete; work order 3 is a no-install synthetic physiology/confound fixture pack. Read docs/SYNTHETIC_MOTOR_FIXTURE_PREREGISTRATION.md and registries/synthetic_motor_fixture_contract.v0.json for its frozen seed 5503, 96-item, eight-family, 48/32/16, 4 MiB fixture-only boundary. Then read docs/SYNTHETIC_MOTOR_FIXTURE_IMPLEMENTATION.md and registries/synthetic_motor_fixture_implementation.v0.json, then docs/SYNTHETIC_MOTOR_FIXTURE_RESULT.md and its result registry. All 18 gates passed after exact-commit green CI: 1.20 seconds, 118,177,792-byte peak RSS, 584,308 output bytes, and zero retained generated files. Work order 3 is complete. Work order 4's no-install classical adapter plan is also complete. No model, training, inference, scoring, S20, PhysioNet, or claim operation is permitted by this result. Read docs/CLASSICAL_EEG_ADAPTER_PREREGISTRATION.md and registries/classical_eeg_adapter_contract.v0.json for the frozen low-frequency shrinkage-LDA, causal CSP-LDA, Riemannian-MDM, grouped-fit, twelve-refusal, no-import boundary. Then read docs/CLASSICAL_EEG_ADAPTER_IMPLEMENTATION.md and its implementation registry, then docs/CLASSICAL_EEG_ADAPTER_RESULT.md and its result registry. Exact implementation eefb7b0 passed CI 31280581308 before one measured roundtrip. All 18 gates passed in 0.12 seconds at 22,822,912-byte peak RSS with a 27,335-byte plan that was removed. Work order 5 is now the synthetic-only contact-mask, noise, and missing-channel contract. It does not authorize hardware, real data, adapter execution, or a scientific claim. Read docs/CONTACT_AWARE_EAR_CHANNEL_PREREGISTRATION.md and registries/contact_aware_ear_channel_contract.v0.json. Seed 5505 freezes 48 items, 16 generic bilateral channels, six separate masks, a fixed maximum-four/minimum-two-per-side target-blind rule, and 16 refusals. Contract commit c6e216f passed CI 31281290300 before implementation. Then read docs/CONTACT_AWARE_EAR_CHANNEL_IMPLEMENTATION.md and registries/contact_aware_ear_channel_implementation.v0.json. The lazy-NumPy fixture, validator, metadata-only inspector, hashes, resource guards, refusal matrix, and CLI were locally qualified. Exact implementation 76ccc63 passed CI 31282344300 before one measured roundtrip. Read docs/CONTACT_AWARE_EAR_CHANNEL_RESULT.md and its result registry. All 18 gates passed in 0.40 seconds at 55,394,304-byte peak RSS with 938,874 output bytes and zero retained generated files. Work order 5 is complete. Work orders 6 and 7 subsequently qualified the strict parser and consumed the one real VHDR compatibility check described below; do not repeat either stage or open a sibling. Then read docs/LOOP_55_PRIMARY_SOURCE_RESEARCH.md and registries/loop55_eeg_neural_effect_research.v0.json. Loop 55 planning separates a higher-power causal performed-hand endpoint from a harder causal 29-key endpoint, makes performed action primary rather than corrected intended text, and keeps the published post-keypress-centered window as a diagnostic only. It recommends a future grouped trial split, twelve matched controls, exact final-trial inference, at most 10,000 parameters and 12 fits, and one-shot final target order. Current local verification passes 24 focused checks plus 9 public-status subtests and the full suite at 1007 passed with 3 expected skips. Loop 55 is still Loop 54 dependent, experimentally Not Started, and unauthorized; no S20 content, split, target, model, training, inference, or score was accessed. Documentation-sync commit 8efcb17 passed push CI 29473032843 and PR #33 CI 29473045583; Base Python and Optional Neuro Readers are green in both. Then read docs/LOOP_55_AI_ASSISTED_REPRESENTATION_RESEARCH.md, registries/loop55_ai_research_policy.v0.json, and fixtures/loop55_ai_synthetic_proposal.v0.json. The additive policy lets AI propose only strict synthetic causal-representation recipes today. The dependency-free validator hashes accepted proposals and rejects target leakage, future context, language models, pretrained weights, unknown fields, protected observations, model/training runs, and resource expansion. Policy commit 8855fae is remotely green; implementation commit bd52cce adds the CLI and 24 focused tests. Historical-hash repair f50be96 passed push CI 29621564301, and the complete local suite passes 1,087 tests with 3 expected skips. This is engineering governance only: no S20 content or scientific result opened, and a future real train-inner agent phase still requires Loop 54 closeout, an exact preregistration, and separate Tier C authorization. Then read docs/OPEN_EEG_R_AND_D_STRATEGY_2026-08-06.md and registries/open_eeg_rd_strategy.v0.json. The current evidence review keeps the compact specialist-first path, adds a prospective 23,248,224-byte PhysioNet motor positive control, recommends one public-data-selected classical baseline plus a causal motor-physiology assay, and defers foundation models to a separate public-data lane. It also defines a future local-first aggregate receipt for EEG contributors. This is additive planning only: no public EEG payload, S20 path, target, model, checkpoint, training, inference, score, or upload was accessed or authorized. Work order 8 then froze and separately authorized that tiny acquisition. Read docs/PHYSIONET_MOTOR_ACQUISITION_PREREGISTRATION.md, its contract, the authorization packet and decision, then docs/PHYSIONET_MOTOR_ACQUISITION_IMPLEMENTATION.md and registries/physionet_motor_acquisition_implementation.v0.json. Authorization decision 00b91ed passed both jobs in CI 31344104565 before the standard-library executor was built. The no-stat/no-network default CLI, exact metadata allowlist, no-redirect/no-retry transport, one-pass opaque hashing, atomic promotion, bounded receipts, and 23 adversarial tests are fixture-qualified. No PhysioNet source metadata, local path, or EDF payload was touched during implementation. Exact implementation 92760ce then passed both jobs in CI 31345401581 before the one nine-file acquisition. Read docs/PHYSIONET_MOTOR_ACQUISITION_RESULT.md and its result registry. All 12 gates passed: exactly 23,248,224 payload bytes, nine one-pass hash matches, 50.682373 seconds, 55,181,312-byte peak RSS, and zero parser, event, target, model, training, inference, scoring, retry, rerun, or work-order-9 operations. Work order 8 is complete and consumed. The payload and private receipts remain Git-ignored. The separate 10 GB allowance is future headroom, not an expansion of this invocation. Work order 9's real execution remains separately gated. Then read docs/PHYSIONET_MOTOR_POSITIVE_CONTROL_PRIMARY_SOURCE_RESEARCH.md, docs/PHYSIONET_MOTOR_POSITIVE_CONTROL_PREREGISTRATION.md, and registries/physionet_motor_positive_control_contract.v0.json. Work order 9 is now prospectively frozen as a prediction + physiology + confound conjunction over only the existing nine EDFs. Runs 03/07 are the only family-selection rows; all run-11 primary and control predictions must freeze in a pushed remotely green hash-only ledger before the 45 final targets open once. The maximum verdict is a three-person motor-compatible public EEG pilot, not brain-specific, typing, language, or generalization proof. Registration itself authorized no operation; the later exact Tier C decision and generated-fixture implementation preserve those boundaries. Registration 3c00557 passed both jobs in CI 31346882592. Read docs/PHYSIONET_MOTOR_POSITIVE_CONTROL_AUTHORIZATION_PACKET.md and registries/physionet_motor_positive_control_authorization_request.v0.json. Request c62b10a passed both jobs in CI 31347209691. The maintainer then supplied the exact sentence, now recorded additively in docs/PHYSIONET_MOTOR_POSITIVE_CONTROL_AUTHORIZATION_DECISION.md and its decision registry. Authorization-only commit da9399c passed both jobs in CI 31348287824 before implementation. Now read docs/PHYSIONET_MOTOR_POSITIVE_CONTROL_IMPLEMENTATION.md and registries/physionet_motor_positive_control_implementation.v0.json. The dry-run-first reader, exact split, target firewall, causal preprocessing, fixed classical families, 12 controls and hashes, isolated scorer, and resource guards passed the generated 135-event roundtrip with zero real or network access. Exact implementation 52b9b15 then passed both jobs in CI 31351728650 before the one target-blind execution. Read docs/PHYSIONET_MOTOR_POSITIVE_CONTROL_PREDICTION_FREEZE.md and its public aggregate ledger. All nine EDF hashes/parses, 135 events, 33 fits, 45 target-blind inferences, and 12 prediction-set hashes completed within the frozen caps. Freeze 01eeff6 passed both jobs in CI 31352250838 before one target delivery and score. Now read docs/PHYSIONET_MOTOR_POSITIVE_CONTROL_RESULT.md and its aggregate result. The selected 8-30 Hz primary failed at 27/45, 0.604 balanced accuracy, and p=0.137, routing WO9-V1. The prespecified 0.5-4 Hz comparator reached 36/45, 0.800 balanced accuracy, and p=0.000183, a strong held-out task- information result that remains cue/confound-compatible because motor physiology and central-over-proxy localization failed. Work order 9 is complete and consumed. Do not reopen, rerun, tune, or promote it into a brain-specific, typing, language, real-time, or hardware claim. Then read docs/PHYSIONET_LOW_FREQUENCY_COHORT_CONFIRMATION_PRIMARY_SOURCE_RESEARCH.md and registries/physionet_low_frequency_cohort_confirmation_research.v0.json. This additive WO9R Tier A design selects twelve untouched participants, S004-S015, and pairs execution 03/07 -> 11 with imagery 04/08 -> 12. It carries the exact fixed 0.5-4 Hz shrinkage-LDA template forward without model selection and freezes participant-level confirmation, execution/imagery transfer, central localization, ocular/visual proxy, early/pre-cue, timing, no-signal, derangement, and displacement questions. The maximum future route is a within-dataset motor-compatible task effect, not brain-specific proof. Then read docs/PHYSIONET_LOW_FREQUENCY_COHORT_CONFIRMATION_PREREGISTRATION.md and registries/physionet_low_frequency_cohort_confirmation_contract.v0.json. Registration 716e543 passed both jobs in CI 31354565966 and binds the exact 72-file, 184,252,032-byte inventory, all model/control counts, the combined target firewall, and one-shot resources. Its metadata pass made no EDF URL request. Finally read docs/PHYSIONET_LOW_FREQUENCY_COHORT_CONFIRMATION_AUTHORIZATION_PACKET.md and its matching request registry. Request 580708f passed both jobs in CI 31355270896. Then read docs/PHYSIONET_LOW_FREQUENCY_COHORT_CONFIRMATION_AUTHORIZATION_DECISION.md and its machine decision. It records the maintainer's actual short-form instruction to stop repeating authorization boilerplate and continue the sole active WO9R packet; it does not pretend the long packet sentence was typed. Decision 1efeac7 is now remotely green in CI 31355944651. Then read docs/PHYSIONET_LOW_FREQUENCY_COHORT_CONFIRMATION_IMPLEMENTATION.md and its machine registry. The standard-library acquisition, compact causal derivative, 144-fit/216-prediction-set target-blind path, combined hash freeze, and isolated scorer are generated-fixture qualified. The measured 72-run fixture used 12.083017 seconds, 260,784,128-byte peak RSS, and 4,215,687 bytes with zero real or network access; its files were removed. Implementation 8242674 passed both jobs in CI 31359548779 before the one real acquisition and target-blind run. Now read docs/PHYSIONET_LOW_FREQUENCY_COHORT_CONFIRMATION_PREDICTION_FREEZE.md. All 72 official hashes matched over 184,252,032 bytes, and the analysis froze 216 prediction sets after 144 fits while exposing zero final targets to the model stage. Freeze 8cd45d7 then passed both jobs in CI 31360781199 before the same 360 targets opened once. Now read docs/PHYSIONET_LOW_FREQUENCY_COHORT_CONFIRMATION_RESULT.md and its result registry. Execution passed at 123/180, pooled balanced accuracy 0.680975, and p=0.002930; imagery passed at 131/180, 0.728014, and p=0.000244. Both cross-task directions were positive. The router stopped at WO9R-R3 because central-over-proxy localization, physiology, and mandatory cue/ frontal controls failed. Preserve the positive task-information result and the equally important brain-specific limitation together. WO9R is consumed; do not rerun, retune, or reopen its private artifacts. Then read docs/IACKD_CUE_ACTION_DISSOCIATION_PRIMARY_SOURCE_RESEARCH.md, registries/iackd_openneuro_metadata_inventory.v0.json, and registries/iackd_cue_action_dissociation_research.v0.json. IACKD-1 turns WO9R's cue-versus-action ambiguity into a reversal: congruent earlier runs fit one fixed low-frequency model, while held-out incongruent runs force measured hand direction and visual target direction to be opposites. Then read docs/IACKD_CUE_ACTION_DISSOCIATION_PREREGISTRATION.md and its contract. Registration e42b799 passed Base Python job 93493810963 and Optional Neuro Readers job 93493811025 in CI 31400450392. It freezes 1,340 raw-source objects and 7,249,113,684 bytes, 30 participant-hand units, a 30 ms motion guard, HEOG/VEOG controls, 300 fits, 420 prediction sets, one combined target delivery, and an IACKD-R0 through IACKD-R4 router. The Request ef78c06 passed both jobs in CI 31401738032; the maintainer's actual packet-bound continuation is recorded in decision 1f48b30, which passed Base Python job 93502398308 and Optional Neuro Readers job 93502398753 in CI 31403012709. Now read docs/IACKD_CUE_ACTION_DISSOCIATION_IMPLEMENTATION.md and its registry. The standard-library acquisition, one-run-at-a-time BrainVision reader, target firewall, 300-fit/420-prediction path, strict aggregate freeze, and isolated scorer were generated-fixture qualified. Exact implementation f5c36ba passed both jobs in CI 31409141349 before real access. Now read docs/IACKD_CUE_ACTION_DISSOCIATION_RESULT.md and its registry. The one acquisition passed all gates over 1,340 objects and 7,249,113,684 bytes. The one analysis then consumed and parked at IACKD-F10: its first lazy BrainVision reader did not satisfy the frozen 32+4 channel gate. It stopped before signal samples, channels TSV, geometry, events, kinematics, targets, models, predictions, freeze, or score. IACKD-1 has no rerun. The retained private bundle is not authorized for post-failure inspection. Next read docs/IACKD_CHANNEL_INVENTORY_PRIMARY_SOURCE_RESEARCH.md, registries/iackd_channel_inventory_research.v0.json, docs/IACKD_CHANNEL_INVENTORY_PREREGISTRATION.md, and registries/iackd_channel_inventory_contract.v0.json, then docs/IACKD_CHANNEL_INVENTORY_IMPLEMENTATION.md and its registry. Public-source review found that the paper does not assert an exact 36-channel file invariant and the authors' two pinned MATLAB pipelines use different deletion vocabularies, including HEO/VEO aliases and TRIGGER. IACKD-H1 freezes a 128-header, 161,792-byte, sibling-blind aggregate audit. Registration 0e52278 passed both jobs in CI 31412667060 before the standard-library parser, strict response boundary, aggregate router, writer, and module CLI were built. One 128-header generated qualification used 0.037819 seconds, 36,634,624-byte peak RSS, and 4,465 output bytes with zero network or real-content access. Exact implementation 16621cc passed Base Python job 93542494819 and Optional Neuro Readers job 93542494839 in CI 31415213841. Next read docs/IACKD_CHANNEL_INVENTORY_AUTHORIZATION_PACKET.md and its all-false request registry. Request 56531c6 passed both jobs in CI 31416489006, and packet-bound decision 04f2706 passed both jobs in CI 31424361969 before the sole audit. Next read docs/IACKD_CHANNEL_INVENTORY_RESULT.md and its registry. All eleven gates passed over 128 headers and 161,792 bytes. Route IACKDH-R5 measured two declarations: 96 have 29 channels without M1/M2, while 32 have 31 channels with M1/M2; all contain HEOG, VEOG, and TRIGGER at 1024 Hz. This explains the consumed reader failure but is not an EEG-signal or neural result. Local bundle, sibling, sample, event, trajectory, target, model, and score counters were zero. IACKD-H1 is consumed with no rerun; a corrected experiment needs a new prospective contract and Tier C gate. Then read docs/IACKD_ROLE_AWARE_DUAL_REVERSAL_RESEARCH.md and its registry. A target-free code audit shows that changing only the failed count would still treat TRIGGER as EEG and retain an invalid EEG-count contract. Next read docs/IACKD_CHANNEL_ROLE_GEOMETRY_PREREGISTRATION.md and its registry. IACKD-H2 is now frozen prospectively over exactly 316 public BIDS role-and-geometry objects and 457,602 bytes, with no local-bundle access. It fixes source-declared channel roles, explicit unavailable reference and geometry states, aggregate-only output, and an ordered compatibility router. Its generated-fixture implementation may begin only after this registration is committed, pushed, and both CI jobs are green. A separate all-false Tier C packet and fresh decision are still required before any public body is fetched. Only a completed H2 aggregate may freeze IACKD-2's role-map hash and symmetric reversal arms. No H2 body, signal, target, model, or score is open. Registration 228ccd0 passed both jobs in CI 31427931578 before the generated-fixture implementation began. Then read docs/IACKD_CHANNEL_ROLE_GEOMETRY_IMPLEMENTATION.md and its registry. The standard-library module now enforces strict role, sidecar, geometry, transport, pairing, aggregate-output, resource, and five-route contracts. Its final 316-body generated qualification used 457,602 bytes, 0.054680 seconds, 34,996,224-byte peak RSS, and 8,282 output bytes. Constructed IACKDR-R4 proves only fixture mechanics. Forty-seven focused, 1,751 base, and 1,822 optional tests pass locally. Next commit, push, and require both CI jobs green for the exact implementation; only then prepare a separate all-false Tier C request. Do not fetch public H2 content from this record. Exact implementation 9f6fef9 passed both jobs in CI 31430151368, and the all-false request is now prepared in docs/IACKD_CHANNEL_ROLE_GEOMETRY_AUTHORIZATION_PACKET.md and its registry. It binds one possible later 316-request/457,602-byte public metadata audit, but authorizes zero requests now. First commit, push, and remotely green the packet. Then identify that sole packet, exact commit, CI, scope, and gate to the maintainer. Only a fresh unambiguous response after that identification may become a separate decision record; earlier continuation is not retroactive authorization. That fresh response is now continue :), recorded verbatim in docs/IACKD_CHANNEL_ROLE_GEOMETRY_AUTHORIZATION_DECISION.md and its registry. The decision binds request 86174bc and green CI 31431064259 without expanding scope. It is not effective until its own commit is pushed and both CI jobs pass; no H2 body may be fetched before that final gate. Decision f6eb5ab passed both jobs in CI 31444154297, then the sole audit completed and consumed all 316 bodies. Read docs/IACKD_CHANNEL_ROLE_GEOMETRY_RESULT.md and its registry. Route IACKDR-R1 records a frozen control-taxonomy mismatch: HEOG, VEOG, and Trigger are all source-typed MISC, while count, sampling, 26-channel core, average reference, and geometry are stable. There is no rerun or admissible H2 role-map hash, and the retained bundle remains closed. Then read docs/IACKD_SOURCE_DECLARED_CONTROL_POLICY_RESEARCH.md and its registry. H3 is a separately named artifact-only repair that preserves BIDS 1.7.0 source counts before assigning functional roles. Its candidate policy is not a validated reader. Research ed5ce82 is remotely green; now read docs/IACKD_SOURCE_SEMANTICS_IMPLEMENTATION.md and its registry. The module has only plan, generated-fixture, and inspect modes. Then read docs/IACKD_SOURCE_SEMANTICS_RESULT.md: implementation 8c5784a was green before one measured generated closeout passed all gates. H3 validates policy mechanics, not a real reader or neural result; real data and IACKD-2 stay closed. Next read docs/IACKD_ROLE_AWARE_DUAL_REVERSAL_PREREGISTRATION.md and its machine contract. The prospective IACKD-2 design binds H3's fixed 26-channel semantics and requires both opposing mapping transfers to favor action over the exact cue surrogate. It freezes 660 fits, 900 prediction sets, a participant-level weaker-arm statistic, central/occipital/EOG/timing controls, one remote-green freeze, and one target delivery and score. Its storage-safe future path streams one registered run group at a time under a 1 GiB disk cap and forbids the existing private bundle. Registration 5bdab30 passed both jobs in CI 31448911258; then read docs/IACKD_ROLE_AWARE_DUAL_REVERSAL_IMPLEMENTATION.md and its registry. The generated-only implementation separates model and scorer objects, exercises the exact 660-fit/900-prediction matrix twice, and recomputes the freeze before generated target delivery. Portability correction af7488a passed both jobs in CI 31451262840 before one registered generated closeout passed all 15 gates in 5.024801 seconds at 257,130,496-byte peak RSS with 30,170 output bytes. Read docs/IACKD_ROLE_AWARE_DUAL_REVERSAL_SYNTHETIC_RESULT.md and its registry. Constructed IACKD2-R5 has no scientific value; the closeout is consumed with no rerun. Then read docs/IACKD_ROLE_AWARE_DUAL_REVERSAL_AUTHORIZATION_PACKET.md and its machine request. The all-false packet freezes one fresh 1,340-object stream, one-run-at-a-time derivatives, 660 fits, 900 target-blind predictions, one green freeze, one target delivery, and one score. It authorizes nothing. Commit, push, and require both CI jobs green before identifying it as the sole Tier C gate; only a fresh later maintainer continue may bind it. Request 862141f then passed both jobs in CI 31454131606, the packet was identified as the sole active Tier C gate, and the maintainer's fresh continue was recorded verbatim. Read docs/IACKD_ROLE_AWARE_DUAL_REVERSAL_AUTHORIZATION_DECISION.md and its registry. Decision 2ce87fa passed both jobs in CI 31456317734 before the real executor was built. Now read docs/IACKD_ROLE_AWARE_DUAL_REVERSAL_REAL_IMPLEMENTATION.md and its registry. The exact executor streams one registered run group at a time, validates 29/31 source semantics and available geometry, creates separate target-free model, sealed scorer, and physiology derivatives, runs exactly 660 fits and 900 predictions, and refuses score access until its aggregate freeze is remotely green. One generated qualification passed all 15 gates in 5.604450 seconds at 270,745,600-byte peak RSS with 4,523 output bytes and zero public, real-data, or old-bundle operations. This is still engineering evidence only. Next commit and remotely green this exact implementation; only then run the one fresh public stream. Never use the retained bundle. Exact implementation dab5dd4 passed both jobs in CI 31461818620 before the sole stream. Now read docs/IACKD_ROLE_AWARE_DUAL_REVERSAL_STREAM_RESULT.md and its registry. The consumed invocation opened only the first registered metadata response and parked at IACKD2-F08 because its Content-Length did not satisfy the exact 1,178-byte gate. The actual header value is unavailable; no body was read, so the result does not establish whether content changed. No selected object, EEG signal, target, derivative, model, freeze, delivery, or score was reached. IACKD-2 has no retry or rerun, and its downstream analysis is unreachable. Any transport-stable recovery must be a separately named prospective lane with fresh authorization. Then read docs/IACKD_TRANSPORT_STABLE_RECOVERY_RESEARCH.md, registries/iackd_transport_stable_recovery_research.v0.json, docs/IACKD_TRANSPORT_STABLE_RECOVERY_PREREGISTRATION.md, and registries/iackd_transport_stable_recovery_contract.v0.json. This new IACKD-T1 registration changes only small-metadata framing: exact observed bytes and SHA-256 remain authoritative, while valid fixed-length, chunked, or close-delimited transport can be recorded. Large selected objects and every IACKD-2 scientific field remain unchanged. No public request or real executor is authorized. Registration ee0f62a passed both jobs in CI 31472269070; now read docs/IACKD_TRANSPORT_STABLE_RECOVERY_IMPLEMENTATION.md and registries/iackd_transport_stable_recovery_implementation.v0.json. The dependency-free module has no URL opener or execute mode, and its one final generated closeout passed 10 acceptance validations and 22 refusals with zero real/public/model operations. Exact implementation 93a067c passed Base Python job 93724709807 and Optional Neuro Readers job 93724709840 in CI 31474412246. Now read docs/IACKD_TRANSPORT_STABLE_RECOVERY_AUTHORIZATION_PACKET.md and registries/iackd_transport_stable_recovery_authorization_request.v0.json. The all-false IACKD-2R request binds one future additive executor, one fresh stream, the unchanged target-blind matrix, one green freeze, and one score, with a new pre-consumption machine-load gate. It authorizes nothing. Commit, push, and require both CI jobs green, then stop and identify it as the sole Tier C packet. Request 525e97e passed both jobs in CI 31475356506; the maintainer then supplied a fresh continue. Read docs/IACKD_TRANSPORT_STABLE_RECOVERY_AUTHORIZATION_DECISION.md and its registry. This additive record quotes the actual word and binds the packet without expansion. Decision feef8f7 passed Base Python job 93730242015 and Optional Neuro Readers job 93730242090 in CI 31476158747. Now read docs/IACKD_TRANSPORT_STABLE_DUAL_REVERSAL_REAL_IMPLEMENTATION.md and registries/iackd_transport_stable_dual_reversal_real_implementation.v0.json. The additive module has a new isolated root and no consumed-executor import, call, or old-root argument. Its generated qualification passed 18/18 gates, all three metadata framing profiles, 660 fits, 900 predictions, deterministic replay, and 13 mutations in 4.939357 seconds at 261,488,640-byte peak RSS with 5,825 output bytes and zero real/public operations. It also refuses before consumption unless at least 10 GiB is free and normalized one-minute load is at most 1.0. Exact implementation b32dc25 passed both jobs in CI 31478167292 before the sole invocation. Now read docs/IACKD_TRANSPORT_STABLE_DUAL_REVERSAL_STREAM_RESULT.md and its registry. The machine gate passed and the new marker consumed IACKD-2R. The first metadata body passed status, URL, framing, encoding, and its exact 1,178-byte count, then failed the pinned SHA-256 before parsing. One body was read and hashed; every selected-payload, EEG, event, trajectory, target, derivative, model, freeze, delivery, and score counter stayed zero. IACKD-2R is parked at IACKD2R-F05 with no retry or rerun. Do not fetch or parse the changed body under this contract. Then read docs/IACKD_SNAPSHOT_IDENTITY_RECOVERY_RESEARCH.md and registries/iackd_snapshot_identity_recovery_research.v0.json. IACKD-M1 replaces the mutable root-body identity assumption prospectively with four independent gates: snapshot hexsha, recursive content-addressed file tree, selected acquisition inventory, and critical Name/BIDSVersion/License/DOI compatibility. Official pinned OpenNeuro source shows that public recursive file rows include full paths, Git object IDs, sizes, annexed status, and S3 versionId URLs. No dataset-specific API response or local IACKD path was accessed. Next implement only the generated canonicalizer after a frozen contract; one later 2 MiB public metadata response remains separate Tier C work. Never treat this research or an earlier continue as permission. Research 723c8e2 passed both jobs in CI 31480538821. Now read docs/IACKD_SNAPSHOT_IDENTITY_PREREGISTRATION.md and registries/iackd_snapshot_identity_contract.v0.json. The exact query, request hash, 1,679-row tree, 1,340-row selection, twelve role summaries, critical metadata, private-manifest boundary, refusal matrix, and resource caps are frozen. Registration 1667e30 passed both jobs in CI 31481270697. Then read docs/IACKD_SNAPSHOT_IDENTITY_IMPLEMENTATION.md and its registry. The generated-only standard-library canonicalizer now passes 49 focused, 2,084 base, and 2,155 optional tests. Its final constructed roundtrip passed all 37 refusals and two replays in 0.888773 seconds at 38,436,864-byte peak RSS with 426,792 output bytes. It has no network client, execute mode, real endpoint, or local IACKD path. Next commit, push, and green this exact implementation. Only afterward prepare an all-false Tier C request for one public response; no public API, EEG payload, target, model, or score access is currently open. Exact implementation 7b8f47b passed both jobs in CI 31483435801. Now read docs/IACKD_SNAPSHOT_IDENTITY_AUTHORIZATION_PACKET.md and registries/iackd_snapshot_identity_authorization_request.v0.json. The request is all false: it describes one future generated/mock transport wrapper and, only after separate green proof, one exact 355-byte GraphQL POST with one 2 MiB-capped response and zero payload requests. Next commit, push, and green this packet, identify it as the sole Tier C request, and stop for a fresh decision. Do not implement the wrapper or make a public request from the packet or an earlier continue. The maintainer then said keep going, move the needle, continue, you approved to go on. Read docs/IACKD_SNAPSHOT_IDENTITY_AUTHORIZATION_DECISION.md and its registry. The decision quotes that exact message and binds only green request ce84738. Decision 4165c24 passed both jobs in CI 31485359989 before wrapper work began. Now read docs/IACKD_SNAPSHOT_IDENTITY_PUBLIC_IMPLEMENTATION.md and its registry. Its standard-library wrapper passed one generated/mock qualification with 20 wrapper refusals, two replays, 429,430 output bytes, zero public requests, and zero neural or target operations. Next commit, push, and green this exact wrapper. Do not make the one public request before both jobs pass, and do not request an EEG payload at all. Wrapper 406bff8 then passed both jobs in CI 31487183289 before the one registered request. Read docs/IACKD_SNAPSHOT_IDENTITY_PUBLIC_RESULT.md and its result registry. The 595,082-byte response failed the exact top-level {data} field-set gate at IACKDMP-F05; the raw body was discarded, no private selected manifest or payload operation followed, and the lane is consumed with no retry or rerun. Do not infer the unknown response field or patch the consumed executor. Then read the CML-v0 architecture research followed by docs/CAUSAL_MOTOR_LATTICE_SYNTHETIC_PREREGISTRATION.md, docs/CAUSAL_MOTOR_LATTICE_SYNTHETIC_IMPLEMENTATION.md, and docs/CAUSAL_MOTOR_LATTICE_SYNTHETIC_RESULT.md with their matching registries. Contract 67709a3 and implementation 90fa467 were each remotely green before one seed-5513 run of the exact 4,535-parameter model. It reached 1.0 hand/key accuracy on 16 constructed signal-bearing check rows, localized all three branches, preserved causal future-tail behavior, and replayed exactly. Eighteen of 19 gates passed. Float32 common-mode logit error 1.9073486e-6 exceeded the frozen 1e-6 tolerance, so the run parked at CML-R0; synthetic final stayed closed and there is no rerun. Runtime was 6.553 seconds at 398,737,408-byte peak RSS with 37,371 output bytes. Public downloads, Loop 54-B/C, S20 modeling, protected targets, hardware, and every scientific claim remain unauthorized. Then read docs/FOUNDATION_MODEL_DECODER_STRATEGY_2026-08-06.md and registries/foundation_model_decoder_strategy.v0.json. The product architecture now treats the compact causal model as a sensor-to-evidence adapter and selects frozen gpt-5.6-sol as the first hosted language decoder candidate. The matched FM-A02 condition must beat both CTC-only and fixed item-deranged evidence under the same model. Then read docs/FOUNDATION_MODEL_BRIDGE_V0.md and registries/foundation_model_bridge_v0.json: FM-0 is implemented as a synthetic no-call compiler. It transformed 7,327 input bytes into 12 plans totaling 34,349 bytes in 0.00275 seconds at about 21.5 MB peak RSS. No provider call, credential, raw neural upload, protected row, target, model, score, or scientific claim opened. Then read docs/FOUNDATION_MODEL_LIVE_SMOKE_PREREGISTRATION.md, docs/FOUNDATION_MODEL_LIVE_SMOKE_AUTHORIZATION_DECISION.md, docs/FOUNDATION_MODEL_LIVE_SMOKE_IMPLEMENTATION.md, and registries/foundation_model_live_smoke_implementation.v0.json, then docs/FOUNDATION_MODEL_LIVE_SMOKE_RESULT.md and registries/foundation_model_live_smoke_result.v0.json. FM-1 is consumed and parked after its one synthetic gpt-5.6-terra invocation: three calls were attempted, two strict responses completed, and the first matched FM-A02 response had non-completed provider status. No FM-A03 or matched comparison exists. Runtime was 8.406004375 seconds at 39,337,984-byte peak RSS. Protected reads, targets, uploads, training, fine-tuning, and scoring stayed zero. Do not rerun or tune from FM-1. It has no target and establishes no decoding or neural value. Then read docs/AI_LOCAL_FIRST_R_AND_D_BUDGET_2026-08-08.md and registries/ai_local_first_rd_budget.v0.json. The user authorized a $50 aggregate AI-provider ceiling. Reserve $0.50 for FM-1 accounting and treat the remaining $49.50 as gated ceilings, not a spending target. Prefer local MNE, MOABB, pyRiemann, and compact Braindecode work. The Apple earbud electrode application is a sensor-architecture lead, not evidence that current AirPods provide EEG or thought decoding. No hardware, protected data, target, large download, or claim permission follows from this budget. Then read docs/LOOP_56_PRIMARY_SOURCE_RESEARCH.md and registries/loop56_cross_modality_accessibility_research.v0.json. Loop 56 freezes five verdict classes, a 12-level capability ladder, 18 comparison dimensions, 16 claim fields, and a 12-part at-home conjunction. It separates continuous input from causal incremental output and measured end-to-end latency. Current local MEG and historical EEG tiny-model evidence is negative, not matched; fresh EEG evidence is unavailable. Planning is complete, but the final cross-modality verdict is Not Started, Loop 55 result dependent, and separately unauthorized. Planning commit 6583ca3 passed push CI 29586877054 and PR #34 CI 29586915269; Base Python and Optional Neuro Readers are green in both.
Read docs/LOOP_30_PRIMARY_SOURCE_RESEARCH.md for the future loopback-only target-free replay inspector, nine-clock timing contract, six latency claim levels, 18 future gates, and 30 refusals. It is planning research only; Loop 30 remains Not Started with no trace, UI, server, or browser run.
Read docs/LOOP_31_PRIMARY_SOURCE_RESEARCH.md for the attribution taxonomy. Its ten-condition encoder slice ran inside the consumed shared event: zero-signal and timing-only components passed, but the complete registered conjunction failed. Loop 31 is parked; sensor-signal dependence and brain-specific origin were not established, and the five-condition LLM/Neuro Token extension remained closed.
Read docs/LOOP_35_PRIMARY_SOURCE_RESEARCH.md for the ten confound classes, nine future synchronized stream classes, 13 conditions, three-stage access boundary, and the claim ceiling of incremental brain-sensor information beyond recorded controls. It is planning research only; Loop 35 remains Not Started and unauthorized.
Read docs/LOOP_36_PRIMARY_SOURCE_RESEARCH.md for the six representation layers, five modality profiles, 24 channel fields, 12 operation classes, explicit unit/transform/reference/interpolation boundaries, 22 gates, and 30 refusals. It is planning research only; Loop 36 remains Not Started, unauthorized, and capped at future declared metadata compatibility.
Read docs/LOOP_37_PRIMARY_SOURCE_RESEARCH.md for the six export layers, five artifact profiles, stable BIDS 1.11.1 field mapping, explicitly non-standard NeuroDecodeKit payload boundary, 24 gates, and 32 refusals. It is planning research only; Loop 37 remains Not Started and unauthorized.
Read docs/LOOP_38_PRIMARY_SOURCE_RESEARCH.md for the five sensitivity levels, eight artifact classes, ten lifecycle surfaces, 12 sensitive-field classes, five deletion-receipt levels, 26 gates, and 36 refusals. It is planning research only; Loop 38 remains Not Started, unresolved copies remain unresolved, and all scanner/deletion/identity/release work is unauthorized.
Read docs/LOOP_39_PRIMARY_SOURCE_RESEARCH.md for the seven qualification levels, 18 environment fields, eight output classes, six comparison classes, six future matrix cells, 28 gates, and 38 refusals. It is planning research only; Loop 39 remains Not Started, Python 3.10/macOS/cross-OS support stays unqualified, and all fixture/manifest/matrix/install/build work is unauthorized.
Read docs/LOOP_41_PRIMARY_SOURCE_RESEARCH.md for the six-layer future RW3-to-NeuroToken join, seven-clock ledger, anomaly and resume rules, and explicit dependency firewall. It is planning research only; Loop 41 remains Not Started and unauthorized, with no fixture or runtime.
Read docs/LOOP_42_PRIMARY_SOURCE_RESEARCH.md for the Q0-only OpenBCI Cyton base 8-channel USB-radio selection, 28-field identity record, packet/ timing/locality/safety boundaries, four stages, 34 gates, and 46 refusals. It is planning research only; no purchase, SDK, connection, participant, recording, device qualification, or scientific result is authorized.
Review docs/RW3_STAGE_A_AUTHORIZATION_PACKET.md only when deciding whether to authorize the independent RW3 Stage A.
Read docs/NEXT_20_LOOPS_PRIMARY_SOURCE_RESEARCH.md and docs/LOOPS_25_44_ROADMAP.md for the planning-only Loops 25-44 queue.
Paste prompts/CODEX_START_PROMPT.md into Codex to continue.
Run tests:

PYTHONPATH=src python -m unittest discover -s tests -v

Tracker scope: the spreadsheet's 20 KPI is the size of the original roadmap, not the current loop number. The workbook remains the last visually reviewed Stage A snapshot because its prior import reached 1.57 GiB peak RSS. The current sources of truth are docs/NEXT_20_LOOPS_TRACKER.md and registries/next_scientific_loops.v0.json; they report Loops 26/31/33 consumed and parked; Loop 48 Stage A F5, consumed Stage B H4 diagnostic, and consumed failed Stage C synthetic gate, all with no rerun. The workbook's ninth sheet contains the separate Loops 25-44 queue. Loop 25 passed once at its target-free mechanics boundary and now has execution_authorized: false because no rerun is open; Loop 26 was preregistered at 881145d, authorized at 1c0e52c, and scored once after remote-green prediction freeze 54bdca9. Its candidate lost to the prior and the event is consumed; Loop 27 planning research selects S25 metadata while preregistration remains blocked; Loop 28 planning research defines a strict zero-shot final-only rule while its experiment remains Not Started; Loop 29 planning research defines separate scalp EEG and OPM-MEG paths while its experiment remains Not Started; Loop 30 planning research defines a target-free local replay boundary while its experiment remains Not Started; Loop 31's 10-condition encoder slice failed its complete shared attribution conjunction while its contingent LLM/Neuro Token extension remained closed; Loop 32 planning research defines the causal 32-parameter fresh-person calibration boundary while its experiment remains Not Started; Loop 33 planning research defines the bounded 8, 16, 24, 32, 44, 55 unique-sentence curve and shared-validation order; it ran once and failed at the 55-row-versus-prior rule, so no rerun or scaling claim is authorized; Loop 34 planning research defines the fresh three-way confidence firewall while its experiment remains Not Started and confidence is unavailable; Loop 35 planning research defines the staged peripheral-confound firewall while its experiment remains Not Started; Loop 36 planning research defines the geometry/reference identity firewall while its experiment remains Not Started; Loop 37 planning research defines the BIDS-envelope/non-standard-payload firewall while its experiment remains Not Started; Loop 38 planning research defines the privacy/lifecycle and deletion-claim firewall while its experiment remains Not Started; Loop 39 planning research defines the cross-machine environment, semantic-identity, and numerical-tolerance firewall while its experiment remains Not Started; Loop 40 planning research defines the edge-package, host-state, fallback, complete-cost, and target-identity firewall while its experiment remains Not Started; Loop 41 planning research defines the stream-to-NeuroToken clock, anomaly, state, schedule, and provenance firewall while its experiment remains Not Started and unauthorized; Loop 42 planning research selects OpenBCI Cyton base 8-channel USB-radio for future mechanics only while its experiment remains Not Started and unauthorized; Loop 43 planning research defines the independent artifact-reproduction firewall while its challenge remains Not Started and unauthorized. Loop 44 artifact-only claim review is complete; engineering release is held and scientific performance release is parked. The maximum future local Loop 31 claim is sensor-signal dependence, and brain-specific attribution remains blocked on Loop 35. All are unauthorized. The queue does not reopen Loop 24 or authorize RW3.

The parallel Real-World Practice Track starts at docs/REAL_WORLD_PRACTICE_TRACK_RESEARCH.md. RW0 is a completed research gate; RW1 is closed as a synthetic-fixture metadata-only interface gate. RW2 is now closed at exact synthetic compatibility level 2: 38 generated recordings are readable and two malformed/unsafe layouts refuse exactly across BrainVision, EDF/EDF+, BDF, EEGLAB external-FDT, FIF, and BIDS. RW3's offline replay/live-source protocol is now frozen at commit c3d1f01; it is a registration result, not a runtime result. Commit 163ff2f adds a hash-bound Stage A decision packet whose machine request keeps authorized_now set to false. Loop 24 ran once under its frozen contract after separate authorization and implementation commits were pushed and remotely green. Float16 preserved behavior but was slower; QNNPACK qint8 reduced payload but failed behavior and runtime gates; no candidate qualified, seed 2402 stayed unopened, and the gate parked at 65.154951 seconds versus its 60-second cap. Seed 2401 is consumed and must not be used for tuning. No real data, targets, training, energy measurement, RW3 Stage A, BrainFlow/LSL/PyXDF, sockets, streams, hardware, consumed caches, S20, and automatic cleaning remain unauthorized. Loop 24 and RW3 remain independent. Loop 25's original registration at a36d97b remains immutable history. The green v1 amendment b6b92d8, authorization commit 1e7296a, and implementation commit 439f151 preceded one registered run. The static 65,537-point response and 23-probe alias gate passed before either partition opened; development seed 2501 then passed and froze, and qualification seed 2502 passed once unchanged. All 24 target-free items, 168 schedule checks, 240 resume checks, and 72 future- mutation controls passed with zero protected reads. Static plus complete-gate runtime was 5.542175 seconds, maximum RSS was 136,806,400 bytes, and generated artifacts totaled 788,967 bytes. The producer is causal with zero right context, but retained neural information and end-to-end latency remain unmeasured. No rerun is authorized. Loop 26 planning at 03605c5 has advanced to a green shared preregistration at 881145d: a 2,908-parameter causal candidate, a 2,884-parameter linear comparator, 21 fits, 24 target-blind inferences, six priors, 31 frozen prediction sets, the complete encoder-attribution matrix, the bounded 8, 16, 24, 32, 44, 55 curve, and all 64 paired sign assignments over six validation sentences. The request remains unauthorized. No real-cache value, target delivery, model run, training run, prediction, or validation score occurred. The next decision is the exact sentence in docs/LOOP_26_AUTHORIZATION_PACKET.md or hold. Loop 29 planning research records scalp EEG as the immediate local-first lane, OPM-MEG as the same-modality partner/lab lane, and cryogenic MEG as the scientific reference. Its 15 requirements, four modality profiles, six qualification levels, 12 future packet gates, and 24 false authorization flags are machine checked. The preferred 5,000,000,000-byte and absolute 10,000,000,000-byte incremental storage limits are capacity only. S20 plus S25 would total 1,106,030,247 bytes, but no Loop 29 download, real-data read, model, stream, device, partner, or hardware operation occurred. The experiment remains Not Started. Loop 30 planning research freezes four source modes, a 30-field target-free trace, nine clock domains, six latency claim levels, 18 future requirements, 30 refusal IDs, and 30 false authorization flags. A future interface must bind exactly to 127.0.0.1, show revision/finalization/proof labels accessibly, and fail on any non-loopback browser traffic. No trace seed, fixture, UI, server, browser run, consumed artifact, model, stream, live source, or hardware operation exists; the experiment remains Not Started. Loop 31 planning research freezes ten encoder conditions, five contingent language-model conditions, six distinct claim classes, 18 future requirements, 24 refusal IDs, and 19 false authorization flags. The encoder conditions now share the green Loop 26 preregistration; the language-model extension and all execution remain closed. No protected cache, target, checkpoint, model, training, validation, LLM, Neuro Token, S20, S25, stream, or device operation exists. Even a future clean local pass can establish only sensor-signal dependence until Loop 35 excludes peripheral and task-locked shortcuts. Loop 32 planning research recommends one pointwise causal 32-parameter hidden affine adapter, four distinct zero-shot/unlabeled/label-light/supervised modes, the nested 0, 2, 4, 8, 16, 32 sentence schedule, and physically separate 32-row calibration, 16-row selection, and 48-row final floors. Its 20 future requirements, 26 refusals, and 22 false authorization flags are machine checked. The experiment remains Not Started; no candidate or mode is selected, S25 remains final-only, and participant/cache/signal/label/target, checkpoint/model, adapter-fit, training, evaluation, download, stream, device, and hardware operations remain unauthorized. Loop 33 planning research recommends six strictly nested unique-sentence prefixes of 8, 16, 24, 32, 44, 55, at most three optimization seeds and 18 candidate fits, size-matched no-signal priors, and a single target-blind shared Loop 26/31/33 validation event. The bounded curve now shares the green Loop 26 preregistration, but execution remains unauthorized. The current rows support no physical- repetition comparison, additional acquisition is not recommended, and all protected/model/training/scoring/device work remains unauthorized. Loop 34 planning research separates seven confidence semantics, eight score and control roles, and recommended fresh synthetic 128/64/256 calibration/selection/final counts. The six real validation rows stay reserved for the shared Loop 26/31/33 event and cannot fit and independently qualify confidence. The experiment remains Not Started; all 26 authorization fields are false, confidence is unavailable, and fixture, confidence-fit, target, scoring, product-confidence, real-data, and download work is unauthorized. Loop 35 planning research freezes ten confound classes, nine future synchronized stream classes, 13 conditions, three stages, 24 gates, 32 refusals, and 31 false authorization fields. Current S21/S7 evidence lacks a fresh complete peripheral- control set. The maximum future local claim is incremental brain-sensor information beyond recorded controls; absolute brain origin, no-keypress transfer, and patient benefit remain unavailable. The experiment is Not Started and unauthorized. Loop 36 planning research freezes six representation layers, five modality profiles, a 24-field future channel record, 12 operation classes, 16 fixture families, 22 gates, 30 refusals, and 29 false authorization fields. A future real-header pass can establish at most declared metadata compatibility. Numerical compatibility, rereference/interpolation behavior, model transfer, and device equivalence remain unavailable. The experiment is Not Started and unauthorized. Loop 37 planning research freezes six export layers, five artifact profiles, 15 standard BIDS mappings, 16 explicit NeuroDecodeKit extension fields, 20 fixture families, four stages, 24 gates, 32 refusals, and 29 false authorization fields. NeuroToken NPZ caches, split reports, and report cards remain non-standard payloads. The experiment is Not Started and unauthorized; no fixture, exporter, derivative tree, validator, protected payload, raw copy, release, or upload exists. Loop 38 planning research freezes five sensitivity levels, eight artifact classes, ten lifecycle surfaces, 12 sensitive-field classes, 12 threat scenarios, five deletion-receipt levels, 24 fixture families, four stages, 26 gates, 36 refusals, and 32 false authorization fields. NIST PF 1.0 is pinned; PF 1.1 remains draft. Unknown copies remain unresolved, and path absence is not media sanitization. The experiment is Not Started and unauthorized; no fixture, scanner, deletion, protected-root scan, identity attack, history rewrite, consent determination, release, or upload exists. Loop 39 planning research freezes seven qualification levels, 18 environment identity fields, eight output classes, six comparison classes, six required future cells, 20 fixture families, four stages, 28 gates, 38 refusals, and 36 false authorization fields. Current public CI is one floating Ubuntu/Python 3.12 environment; Python 3.10 has an unresolved tomllib collection gap, macOS and cross-OS support are unqualified, and no exact dependency lock or built-package identity exists. The experiment is Not Started and unauthorized; no fixture, manifest, matrix job, install, lockfile, package, protected read, model, training, edge, stream, device, or hardware operation exists. Loop 40 planning research freezes seven qualification levels, six package layers, four unselected backend profiles, 20 identity fields, eight output classes, six comparison classes, 24 fixture families, four stages, 30 gates, 40 refusals, and 40 false authorization fields. The retained 1,130-parameter, 5,210-byte float32 causal producer remains blocked by the unexecuted Loop 39 matrix and the absence of a named target OS/architecture/ABI. ExecuTorch/ XNNPACK is a research lead only beside ONNX Runtime Mobile, LiteRT, and Core ML. The experiment is Not Started and unauthorized; no fixture, install, export, package, inference, profiler, simulator, app, device, or hardware operation exists. See docs/LOOP_40_PRIMARY_SOURCE_RESEARCH.md and registries/loop40_research_boundary.v0.json. The exact four-file, 96,090,264-byte S20 EEG acquisition passed once and is consumed. Read docs/LOOP_53_ACQUISITION_RESULT.md and registries/loop53_acquisition_result.v0.json for the measured pass, zero forbidden counters, private receipt bindings, and claim ceiling. The payload remains uninterpreted: Loop 54 header, marker, signal, MAT, target, geometry, event, and trial work still requires its own exact staged Tier C decisions. The older broad S20 packet remains historical and grants no future permission.

Current verified state: two complete S21 SpanishBCBL MEG recordings, one 94,842,381-byte S7 EEG bundle, and one 96,090,264-byte opaque-verified S20 EEG bundle have been selectively acquired under exact caps; the full dataset and 12.79-GB EEG subtree have not been downloaded. Loops 9-12, 14-22, and 23.5 are complete; Loops 13, 23, and 24 are parked after measured gates. Loop 24 retains float32, consumed selection seed 2401, and left qualification seed 2402 physically unopened. Loop 25 is complete at its one-time target-free mechanics boundary; seeds 2501 and 2502 are consumed for this gate and no rerun is authorized. Loop 26/31/33 now has one green shared preregistration, a separately green authorization at 1c0e52c, and a synthetic-only staged implementation. Protected execution waits for the implementation commit to become remotely green. All other Loops 27-44 remain planning-only or artifact-only at their documented boundaries. Loop 27's planning research is complete: S25 session 2 block 2 is the smallest eligible strict MEG pair at 1,009,939,983 bytes, but no preregistration, acquisition request, download, local MAT hash, header, signal, target, model, or transfer operation is authorized. Loops 28-44 have no authorized execution. Loop 28 planning research is complete: it separates T0- T3 evidence, reserves S25 for zero-calibration T2, and recommends a minimum 48 final rows, 0.05 macro sentence-CER advantage, 65,535 paired random assignments plus observed, and strict corruption-control wins. All 21 authorization fields are false and the experiment remains Not Started. Loop 29 planning research is complete while its experiment remains Not Started. Loop 30 planning research is complete while its target-free replay experiment remains Not Started. Loop 31's encoder-only matrix is frozen inside the shared Loop 26 event, while its language-model/NeuroToken extension remains unpreregistered; no Loop 31 execution exists. Loops 32-44 remain at their documented planning or artifact-only boundaries. Session 1 provides a strict 55/6/5 split with train-only robust scaling. Session 2 is a complete two-part FIFF recording with 63 performed trials; MAT slots 54, 58, and 60 are explicitly empty and are preserved as gaps. Its 102-magnetometer cache is scaled only with frozen session-1 train statistics. The fixed same-subject cross-session tiny CTC fails to generalize: corpus CER 0.9179 versus 0.7755 for the no-signal prior, with a paired interval wholly favoring the prior. There is no reliable neural advantage, unseen-person result, retained-accuracy result, integer-only inference, real-time decoder, or at-home hardware claim. Session 2 is now a consumed evaluation set. Loop 15 Stage B separately proves unlabeled robust channel-affine recovery on a fixed synthetic diagonal shift: identity/adapted holdout CER is 0.3448/0.0000, with zero real cache reads and only 21,354 bytes of artifacts. Loop 16 then tests six calibration sizes, three shift seeds, and three independently calibrated shift families. One 1.26-second synthetic row meets the registered median stationary-diagonal rule, but one seed only ties; the adapter harms all channel-mixing and time-varying holdout seeds. This is not a human calibration-time or real-MEG claim. Loop 17 adds a local artifact-backed evidence console with 19 synthetic examples, aggregate-only real metrics, explicit unavailable confidence, noncausal/task scope, and responsive browser QA. It performs no real model run or raw-data read. Loop 18 adds 11 versioned cards across six exact cohorts and four method families from saved reports only. Four within-cohort rankings are authorized; no global rank exists, all missing recommended metadata stays visible, and the 103,789-byte build opens no cache or holdout. Loop 19 then passes a metadata-only access and dependency gate, lazily aligns all 2,534 MAT triggers to one real S7 EEG file, and writes a 12,428,800-byte 2197 x 61 x 25 cache. Its first transparent within-session classifier is a negative result: 0.91% exact key-label accuracy versus 12.27% for the train-only no-signal prior. This is bridge validation, not useful EEG decoding or a portable-hardware claim. Loop 20 adds a 76,646-byte synthetic NeuroTokenCache v0 shaped 48 x 16 x 32, with timing, masks, modality, geometry availability, strict split/source hashes, and explicit causality/resource boundaries. The deterministic target-free payload replays exactly, and an access-tracked audit now confirms that no source target array is opened. It is not learned, not a decoding result, and not an end-to-end streaming claim. Loop 21 then passes a five-schedule synthetic causal replay gate: zero right context, exact frame/timestamp identity, bitwise stream-schedule identity, 300-byte mutable state, and no target/model/decoder/ real-data access. Transport delay still ranges from 0 to 610 ms depending on chunking, and end-to-end text latency remains unmeasured. Loop 22 then trains one preregistered 1,130-parameter model on synthetic motif frames, selects epoch 34 on validation, opens its eight-item test once, and passes both the signal-free comparison and all five streaming schedules with 300-byte state. That perfect synthetic motif result is a mechanism check, not text or brain decoding; its test is consumed. Loop 23 then implements the preregistered language-model-free greedy and width-8 prefix CTC decoders. Registered validation passes at CER 0.0182 and 7/8 exact, so seed 2303 opens once. Frozen test CER is 0.0545 and all repeated pairs are correct, but exact sequence accuracy is only 5/8 versus the required 6/8. Every error is a complete target plus one false tail symbol; prefix and greedy agree. Loop 23 is parked and its test is consumed. Loop 23.5 then fits one preregistered blank-logit intercept on fresh train-frame labels. Validation reaches 16/16 exact, opening seed 2353 once; the frozen test also reaches 16/16 and CER 0 versus 7/16 and CER 0.0818 without calibration. Nine items are corrected, none worsens, and all resource, access, control, bootstrap, and replay gates pass. This is supervised synthetic calibration only, seed 2353 is now consumed, and Loop 24 requires a separate fresh protocol. That protocol was frozen at 186bb6f, authorized separately, implemented before data access, and executed once. Float16 was exact but slower; qint8 was smaller but incorrect and slower; the total runtime cap failed; seed 2402 stayed unopened; and Loop 24 parked with float32 retained. See docs/LOOP_23_5_BLANK_INTERCEPT_CALIBRATION.md, docs/LOOP_24_PRECISION_RUNTIME_PREREGISTRATION.md, docs/LOOP_24_AUTHORIZATION_DECISION.md, docs/LOOP_24_LOCAL_PRECISION_RUNTIME.md, docs/LOOP_25_PRIMARY_SOURCE_RESEARCH.md, docs/LOOP_25_CAUSAL_PREPROCESSING_PREREGISTRATION.md, docs/LOOP_25_ANTI_ALIAS_AUDIT.md, docs/LOOP_25_CAUSAL_PREPROCESSING_AMENDMENT_1.md, docs/LOOP_25_AUTHORIZATION_PACKET_V1.md, and docs/POST_20_ROADMAP.md.

RW1 adds dependency-free inspect-recording and inspect-intake-report commands for BrainVision, EDF/EDF+, BDF, EEGLAB, FIF, and BIDS level-0 metadata. The 532-byte synthetic roundtrip produced 11,545 bytes of validated JSON/Markdown/audit output in 0.001659 seconds with 21,643,264-byte peak RSS and zero binary/raw/cache/target/model/training/network reads. This is file and report interface proof only. See docs/RW1_METADATA_ONLY_LOCAL_INTAKE.md.

RW2's frozen protocol covers explicit optional MNE adapters for synthetic BrainVision, EDF/EDF+, BDF, EEGLAB external-FDT, FIF, and BIDS fixtures; three-window/32-MiB signal materialization, one-thread, 30-second, 1-GiB RSS, 4-MiB per-run output, and 16-MiB fixture-set caps; strict privacy redaction; descriptive time-domain and Welch PSD summaries; and no source mutation or automatic bad-channel action. The 40-fixture implementation passes with 38 readable fixtures and two exact refusals. One measured FIF roundtrip selects nine channels and three windows, returns 11,520 values, materializes 92,160 bytes, and writes 76,592 report bytes in 3.839168 seconds with 150,749,184-byte peak RSS. Real/cache/target/model/training/network access is zero, source before/after hashes match, producer causality is false, and end-to-end latency is unmeasured. See docs/RW2_SIGNAL_QUALITY_CLOSEOUT.md.

RW3 freezes a future neurodecodekit.source_chunk envelope, separate raw, corrected, and arrival clocks, explicit packet anomalies and reconnect state, five chunk schedules, 18 target-free future fixture families, 30 exact refusal IDs, and four sequential stages. Seven contract tests plus three authorization- binding tests protect that registration and the proposed 90-case Stage A scope. No fixture, payload, source chunk, CLI, optional dependency, board, socket, stream, or XDF file was created or opened. See docs/RW3_REPLAY_LIVE_EQUIVALENCE_PREREGISTRATION.md, registries/replay_equivalence_contract.v0.json, docs/RW3_STAGE_A_AUTHORIZATION_PACKET.md, and registries/rw3_stage_a_authorization_request.v0.json.

Loop 24 executed the float32 eager reference, explicit CPU float16, and dynamic- qint8 QNNPACK candidates around the 1,130-parameter synthetic producer. All 12 balanced rounds and the timing protocol passed. Float16 passed correctness but was 1.170x the producer latency and 1.088x the full-pipeline latency. Qint8 used 47.1% of the float32 numeric payload but changed greedy/prefix behavior and was 2.785x/1.812x the producer/full latency. The 65.154951-second orchestration exceeded the 60-second cap, so Loop 24 is parked with float32 retained and qualification unopened. See docs/LOOP_24_LOCAL_PRECISION_RUNTIME.md.

The current registries keep prompted typing, imagined speech, natural reading, P300, SSVEP, and motor imagery as separate evidence cohorts. They also keep EEG separate from eye tracking, wrist EMG, heart/PPG, IMU, microphone, and hand tracking. No live or portable device is qualified.

Loop 43 Independent Reproduction Boundary

Loop 43 planning research is complete while the challenge remains Not Started and unauthorized. It defines a future commit-reveal reproduction of one target-free NeuroToken causal-replay software artifact with 16 independence fields, 28 packet fields, 34 submission fields, four separately authorized stages, 36 gates, and 48 refusals. No packet, oracle, outreach, contributor, submission, or independent artifact reproduction exists. Even a future pass would be software-artifact reproduction, not scientific replication, neural advantage, decoding, person generalization, or device qualification. Read docs/LOOP_43_PRIMARY_SOURCE_RESEARCH.md before any Loop 43 proposal.

Loop 44 Claim And Release Boundary

Loop 44 artifact-only planning and claim review are complete. Read docs/LOOP_44_PRIMARY_SOURCE_RESEARCH.md, docs/LOOP_44_CLAIM_PROMOTION_AND_RELEASE_DECISION.md, and registries/loop44_claim_release_matrix.v0.json.

The exact decision is an engineering release hold and a scientific performance park. The repository has useful engineering contracts and three retained real-data predictive results, but none demonstrates positive neural advantage. No tag, GitHub release, DOI, participant payload, unseen-person result, end-to-end real-time result, portable/home device result, independent reproduction, or scientific replication exists.

Loops 45-64 Scientific Roadmap

Read docs/LOOPS_45_64_SCIENTIFIC_ROADMAP.md and validate registries/next_scientific_loops.v0.json. The five phases are Real Signal Truth, Unseen-Person Verdict, Accessible EEG Evidence, Causal Local Use, and Independent Evidence And Release. Loop 45 is complete at its target-free mechanics boundary. Loops 46/47 consumed one registered S21 validation event and are parked after the candidate lost to the no-signal prior; no rerun is authorized. Loop 48 completed one post-outcome artifact-only Stage A that selected F5 output-distribution instability as the descriptive phenotype. The result is consumed, root cause remains unresolved, and no rerun is authorized. Authorization commit 5bae880 and implementation commit ca21539 were each pushed and remotely green before the one execution. The additive H1-H6 design then advanced through one exact Stage B: 44 fit rows, 11 target-withheld check rows, 20 fits, 35 target-blind inferences, five priors, 41 frozen prediction sets, and one post-green-freeze check score. The primary candidate lost to the prior (0.953566 versus 0.822045 macro CER), H4 stable nonseparability is supported, and fixed-shift H3 has evidence against it. Because all 55 source-train rows were used by prior Loop 26 fits, the result is E2 pipeline discrimination rather than independent validation or sensor- dependence confirmation. Brain-specific origin remains a separate Loop 35 question. Stage B is consumed with no rerun.

Loop 49 planning research is complete while the experiment remains Not Started. Read docs/LOOP_49_PRIMARY_SOURCE_RESEARCH.md and registries/loop49_research_boundary.v0.json. Pinned public metadata selects S24 session 2 block 2 as the preferred permanently development-only MEG person: one FIF plus one protected MAT log totaling exactly 1,048,579,727 bytes. S24 costs 29,701,559 bytes more than S18 but avoids the published S1/S18 alias and preserves S25 as final-only. A future text-grouped split reserves 16 unique sentence groups for selection and requires at least 32 fit groups, with matching S21 selection text excluded from fit. No S24 local path or payload was opened; the >=48 unique-row, channel, geometry, duration, and overlap gates remain unproven. There is no preregistration, acquisition request, or authorization. Decision 0083 placed any new acquisition after the separately gated Loop 48 Stage B. The consumed result now selects L50-R05, which parks S24 acquisition for this model family.

Loop 50 planning research is complete while the experiment remains Not Started. Read docs/LOOP_50_PRIMARY_SOURCE_RESEARCH.md and validate registries/loop50_research_boundary.v0.json. The design freezes global text grouping, five historical S21 out-of-fold folds, a 16-group S24 development gate, equal-person loss, one shared candidate family, ten controls, an exact 20-update inventory, and a worst-person acceptance rule. All 31 authorization fields are false and no S24, S25, S21 cache, model, training, inference, or score occurred inside that planning snapshot. Stage B has since closed at L50-R05; no S24 operation or Loop 50 execution is authorized. Planning commit 085f341 passed push CI 29458102674 and PR #28 CI 29458116994; both required jobs are green in both workflows.

Loop 53 is Consumed; Acquisition Passed; No Rerun. Loop 54 qualification, Loop 55 neural-effect, and Loop 56 cross-modality accessibility planning research are complete, but their real stages or final verdict remain Not Started and unauthorized. Loops 51-52 and 57-64 remain Not Started. No Loop 54 content interpretation is authorized by the Loop 53 pass. L54-A now has a frozen one-file preregistration and pending authorization packet, but that contract is not content-access permission. Loop 55's synthetic AI proposal guard is implemented, but this does not change the real experiment status or authorize a model operation. S25 remains sealed until a separate final-only zero-shot packet is complete.

Current Next Scientific Lane: MARC-1

Read docs/MARC_1_MULTIMODAL_ARTIFACT_RESOLVED_MOVEMENT_RESEARCH.md, registries/marc1_multimodal_artifact_resolved_movement_research.v0.json, docs/MARC_1_GENERATED_QUALIFICATION_PREREGISTRATION.md, and registries/marc1_generated_qualification_contract.v0.json. Then read docs/MARC_1_GENERATED_QUALIFICATION_IMPLEMENTATION.md, registries/marc1_generated_qualification_implementation.v0.json, docs/MARC_1_GENERATED_QUALIFICATION_RESULT.md, and registries/marc1_generated_qualification_result.v0.json.

WO9R remains the strongest positive result: held-out low-frequency execution and imagery task information across twelve people. Its source interpretation did not pass because early-cue and frontal controls remained strong. MARC-1 does not scale that classifier. It freezes a two-axis falsification design:

Freewill-23 tests self-selected movement timing with EEG, four EOG channels, and synchronized wrist acceleration.
Wrist-45 tests the same causal EEG family against eight EMG channels and synchronized robotic encoder kinematics.

The future top route requires both axes and uses the weaker margin. One-axis success cannot be promoted. The 13.59 GB Freewill ZIP is forbidden as a whole download. The next Tier B gate is now frozen as a dependency-free generated qualification: 14 safe ZIP members, no member-content reads, two modality-role profiles, past-only windows, physical target isolation, twelve comparators, and 24 adversarial refusals. Contract 4494d57 passed both jobs in CI 31502115918. Exact implementation e35a587 then passed both jobs in CI 31505555044 before one generated closeout. That run passed MARC1G-R1 in 0.006589 seconds at 23,511,040-byte peak RSS: 14 members, zero payload-overlap bytes, both modality plans, all twelve comparators, and all 24 refusals passed. The result is consumed and has no scientific value because every real-data, target, model, and score counter remained zero. Read docs/MARC_1_GENERATED_QUALIFICATION_RESULT.md. The next eligible task is the newly specified MARC1-CD1 metadata-only central-directory range audit. Read docs/MARC_1_FREEWILL_CENTRAL_DIRECTORY_RESEARCH.md and registries/marc1_freewill_central_directory_research.v0.json. The design allows a future maximum of three response bodies totaling 17,039,360 bytes: 128 KiB of version metadata, the final 128 KiB of the archive, and at most 16 MiB of central-directory metadata. It requires complete ZIP64 reconciliation inside the tail and has no whole-download or exploratory fallback. This is currently research only: no real metadata response, archive range, or payload is authorized. Research 93faf36 is now remotely green, and the generated/mock contract is frozen in docs/MARC_1_FREEWILL_CENTRAL_DIRECTORY_PREREGISTRATION.md and registries/marc1_freewill_central_directory_contract.v0.json. It fixes a virtual 13.59 GB archive, an 18-entry generated directory, direct and bodyless- redirect transport paths, complete in-tail ZIP64, 32 refusal mutations, and no live endpoint. Contract cf63043 passed Base Python job 93837415016 and Optional Neuro Readers job 93837415174 in CI 31508903399 before the generated-only implementation was written. Read docs/MARC_1_FREEWILL_CENTRAL_DIRECTORY_IMPLEMENTATION.md and registries/marc1_freewill_central_directory_implementation.v0.json. The module has only plan, generated qualify, and aggregate inspect; it has no network opener, real archive path, member reader, or execute mode. It materializes 280,249 generated bytes, validates the 128-KiB tail and 148,910-byte directory, inventories all 18 entries, reconciles complete ZIP64, keeps exact names and offsets private, replays both transport paths, and refuses all 32 mutations. The latest development run used 0.007407 seconds, 26,181,632-byte peak RSS, and 11,573 temporary output bytes, all removed. It had no scientific value. Exact implementation 211fd78 then passed Base Python job 93846584402 and Optional Neuro Readers job 93846584527 in CI 31511626051 before one registered closeout. Read docs/MARC_1_FREEWILL_CENTRAL_DIRECTORY_RESULT.md and registries/marc1_freewill_central_directory_result.v0.json. The run passed MARC1CDG-R1 in 0.006544 seconds at 27,131,904-byte peak RSS, with 280,249 generated input bytes, 11,574 temporary output bytes, 18 entries, 32/32 refusals, 14/14 gates, and zero live, real-data, model, score, or claim operations. Its exact outputs were hash-bound and removed. The generated closeout is consumed with no rerun. A live response remains closed Tier C work. Read docs/MARC_1_FREEWILL_CENTRAL_DIRECTORY_AUTHORIZATION_PACKET.md and registries/marc1_freewill_central_directory_authorization_request.v0.json. The all-false request binds one future generated/mock live-wrapper milestone and, only after that exact wrapper is remotely green, one metadata body, one exact 128-KiB tail, and one conditional directory body under a 17,039,360- byte cap. Request 950796d passed Base Python job 93853089748 and Optional Neuro Readers job 93853089786 in CI 31513578445, after which Codex named it as the sole active Tier C gate. The maintainer's fresh words keep going, move the needle, continue, you approved to go on are now quoted verbatim in docs/MARC_1_FREEWILL_CENTRAL_DIRECTORY_AUTHORIZATION_DECISION.md and its machine registry. That decision remains ineffective until its own exact commit is pushed and both CI jobs are green. Before that proof, do not implement the live wrapper, request a public byte, or create a consumed marker.

Decision 624cc4e99a4aa600b68a333c1bcd84e6cebb9dcd then passed Base Python job 93871192638 and Optional Neuro Readers job 93871192713 in CI 31519016891 before wrapper work began. Read docs/MARC_1_FREEWILL_CENTRAL_DIRECTORY_LIVE_IMPLEMENTATION.md and registries/marc1_freewill_central_directory_live_implementation.v0.json. The additive standard-library wrapper now qualifies exact proof loading, machine gating, bounded manual redirects, three-body transport, reuse of the green ZIP64 parser, private exact inventory, aggregate output, and consumed failure behavior using generated fixtures and mocked HTTP only. One fresh python -S closeout passed MARC1CDL-G1 in 0.006050 seconds at 40,763,392- byte peak RSS with 14/14 gates, 32 inherited and 8 wrapper refusals, 12,182 output bytes, and zero network or forbidden counters. Its temporary files were removed. Exact wrapper 5dfa3c4c8cd7f0e990b7b1db7b35c4df8694171f then passed Base Python job 93879378282 and Optional Neuro Readers job 93879378362 in CI 31521510374 before the one registered live invocation. Read docs/MARC_1_FREEWILL_CENTRAL_DIRECTORY_LIVE_RESULT.md and registries/marc1_freewill_central_directory_live_result.v0.json. MARC1CD-R1 inventoried all 1,227 entries in the current 13,591,548,048-byte public ZIP from 306,758 accepted metadata bytes without downloading the archive or opening a member. All 14 gates passed in 2.727437 seconds at 43,974,656-byte peak RSS. The private exact manifest is Git-ignored; the public result contains only aggregate counts and hashes. MARC1-CD1A is consumed with no retry or rerun. Any member-level selection or access requires a new prospective contract and Tier C decision; no neural or decoding evidence was produced.

That next prospective contract is now docs/MARC_1_PRIVACY_PRESERVING_PILOT_SELECTION_PREREGISTRATION.md with machine registry registries/marc1_privacy_preserving_pilot_selection_contract.v0.json. MARC1-P1 freezes 12 participants per source by DOI-bound SHA-256 rank, 72 Freewill run bundles and 288 opaque core members, 12 Wrist participant archives, a Freewill session-1/session-2 cross-day split, a Wrist runs-1-6/ runs-7-8 split, and an 8-GiB joint ceiling. Selection may not use member size, CRC, event count, target, quality, or outcome information. The current gate is generated implementation only after this exact contract is remotely green; the sealed inventory, Wrist metadata, payloads, signals, targets, models, and scores remain closed.

Contract d1218066e64dea502d263acf0c096ed7eab55a11 passed Base Python job 94028013357 and Optional Neuro Readers job 94028013230 in CI 31569417204 before the generated implementation began. Read docs/MARC_1_PRIVACY_PRESERVING_PILOT_SELECTION_IMPLEMENTATION.md and registries/marc1_privacy_preserving_pilot_selection_implementation.v0.json. The module has only plan, qualify, and aggregate inspect; it builds 1,227 + 55 generated rows, selects the exact frozen cohorts and splits, writes 300 private rows mode 0600, replays under reversed input order, and refuses all 36 mutations. Development MARC1PSG-R1 processed 873,348 bytes in 0.226802 seconds at 32,833,536-byte reported RSS and emitted 182,563 temporary bytes, all removed. It read no real metadata or neural content.

Exact implementation 0c0a6982c6b9c65d6c51413d1baa8b577e00a194 passed Base Python job 94034790262 and Optional Neuro Readers job 94034790315 in CI 31571668853 before the one registered generated closeout. Read docs/MARC_1_PRIVACY_PRESERVING_PILOT_SELECTION_RESULT.md and registries/marc1_privacy_preserving_pilot_selection_result.v0.json. MARC1PSG-R1 passed all 15 gates and 36 refusals in 0.227334 seconds at 32,374,784-byte reported peak RSS. Its 182,564 output bytes were hash-bound, inspected once, and removed. Every real metadata, payload, neural, target, model, score, and claim counter stayed zero. The generated closeout is consumed with no retry or rerun.

Read docs/MARC_1_PRIVACY_PRESERVING_PILOT_SELECTION_AUTHORIZATION_PACKET.md and registries/marc1_privacy_preserving_pilot_selection_authorization_request.v0.json. The all-false MARC1-P1A request freezes a future generated/mock wrapper and, only after that exact wrapper is remotely green, one real metadata selection over a single 418,755-byte no-follow Freewill-manifest read plus one Wrist metadata body capped at 2 MiB. Payload, signal, target, model, score, retry, rerun, and claim operations remain zero and unauthorized.

Request 7f1ba0936e4e0266c0210648aa641feab63cd0eb passed Base Python job 94041819046 and Optional Neuro Readers job 94041819022 in CI 31573969646. After it was identified as the sole active Tier C packet, the maintainer supplied the fresh words approved, continue, achieve a scientific claim, achieve thought to text 😎. Read docs/MARC_1_PRIVACY_PRESERVING_PILOT_SELECTION_AUTHORIZATION_DECISION.md and its machine registry. The decision quotes the actual message, binds the unchanged packet by hash, treats the scientific result as an objective rather than a predeclared outcome, and authorizes zero payload bytes.

Decision 9726d07ab08e9c2815dbe68398659f454693be5e passed Base Python job 94044627592 and Optional Neuro Readers job 94044627647 in CI 31574870204 before implementation. Read docs/MARC_1_PRIVACY_PRESERVING_PILOT_SELECTION_LIVE_IMPLEMENTATION.md, registries/marc1_privacy_preserving_pilot_selection_live_implementation.v0.json, and the two matching live-selector test modules. The additive standard- library wrapper now freezes the exact private source, seven-field Wrist schema, sub-01.zip through sub-45.zip participant rule, known sub-01 anchor, target-field firewall, global-only redirect policy, one-shot marker, machine gate, and private/public outputs. It has no payload, signal, target, model, scorer, override, retry, or fallback interface.

One final generated/mock qualification passed MARC1PSL-G1 with 866,578 input bytes, 214,553 output bytes, 300 private rows, the exact 12+12 cohorts, all 26 refusals, and all 15 gates in 0.183268 seconds at 50,905,088-byte reported peak RSS. All real, network, payload, neural, target, model, score, and claim counters stayed zero; the temporary outputs were removed. Thirty- one focused, 329 MARC, 2,468 dependency-light, and 2,539 optional-neuro tests pass with no new skips.

Exact wrapper 702e61377d41fd1d95939d5e4047be59e4631d4d passed Base Python job 94056321843 and Optional Neuro Readers job 94056321914 in CI 31578614616 before the sole metadata invocation. Read docs/MARC_1_PRIVACY_PRESERVING_PILOT_SELECTION_LIVE_RESULT.md and its machine result. The executor verified the 418,755-byte private inventory exactly once, opened one Wrist response, and failed closed at MARC1PS-F03 because the frozen explicit identity-encoding rule was not met. It read zero public-body or payload bytes, selected zero participants, and performed zero signal, target, model, or score operations. MARC1-P1A is consumed with no retry or rerun. The next prospective gate is a separately named transport- semantics recovery, followed by a new Tier C decision before any real input. This is the same thought-to-text path, not a pivot, and movement evidence is not language evidence.

Read docs/MARC_1_HTTP_IDENTITY_SEMANTICS_RECOVERY_RESEARCH.md and registries/marc1_http_identity_semantics_recovery_research.v0.json for the zero-access MARC1-HT1 diagnosis. RFC 9110 separates request-side identity from response-side applied content codings. The candidate policy accepts an absent response field or one inert identity token and refuses every actual coding. The live header value remains unavailable and must not be inferred. Commit, push, and green this research before freezing a generated-only contract; no implementation or real input is authorized by the research.

Research f515b36cfdd2b297bcbba9885af92e59ead066a7 passed Base Python job 94062432262 and Optional Neuro Readers job 94062432241 in CI 31580575669 before the generated-only MARC1-HT1 contract was frozen. Read docs/MARC_1_HTTP_IDENTITY_SEMANTICS_RECOVERY_PREREGISTRATION.md and registries/marc1_http_identity_semantics_recovery_contract.v0.json. The contract permits only a future zero-network plan/qualify/inspect harness after this exact contract is remotely green. It freezes four accepted forms, 20 refusals, 16 gates, five refusal routes, and exact generated cohort replay. It authorizes no real/private input or payload.

Contract 1f99d0a8c5609dae992fa0e245f179c2f417038f passed Base Python job 94065047494 and Optional Neuro Readers job 94065047277 in CI 31581395690 before the additive MARC1-HT1 implementation was built. Development MARC1HT-G1 passed all 16 gates with four accepted forms, 20 refusals, exact 12+12 cohort replay, 923,052 generated input bytes, and zero real or network bytes. The next gate is implementation commit, push, and both green CI jobs, followed by one registered generated closeout. No live metadata, payload, EEG, target, model, score, or claim action is open. This remains the same thought-to-text research path.

Exact implementation b2cb48cc1c630cf2d22186732e8258619db0a930 passed Base Python job 94073234688 and Optional Neuro Readers job 94073234607 in CI 31583931303 before the registered generated result. The one closeout passed MARC1HT-G1: 4/4 forms, 20/20 refusals, 16/16 gates, 923,052 generated input bytes, 0.112 seconds, 33.1 MB external peak RSS, and zero real/network operations. Its two temporary outputs were removed and no rerun is open. The result must be committed, pushed, and green before an all-false Tier C request may propose a new live metadata wrapper and attempt.

Result 5344d73bb74431e9bba05e3608c2a1523a84cd00 passed Base Python job 94075586323 and Optional Neuro Readers job 94075586171 in CI 31584662864. Read the all-false MARC1-HT1A packet and its machine request. It proposes one additive generated/mock wrapper followed, only after separate green decision and implementation milestones, by one metadata-only selection attempt. Every current authorization flag is false. The sealed upstream 418,755-byte inventory is distinct from the forbidden consumed MARC1-P1A root; no payload, signal, target, model, score, or claim operation is open. MARC1-HT1A is the same path toward a controlled neural positive control and held-out language decoding, not a pivot.

Request 27f39aee5f056eafc81b615cec4a178a41a6c5d2 passed Base Python job 94080678529 and Optional Neuro Readers job 94080678738 in CI 31586256906. Read the separate packet-bound decision. It quotes the maintainer's fresh approval exactly and authorizes the staged wrapper-then-one-attempt sequence only after the decision itself becomes remotely green. Thought-to-text remains the objective, not a predeclared outcome. No wrapper, private read, public request, or payload action has yet occurred under this decision.

Decision 9c7bd48541fbcebabcb9a783cb9047c7f2a2f57a then passed Base Python job 94083644849 and Optional Neuro Readers job 94083644932 in CI 31587195405 before wrapper implementation. Read the additive implementation, its machine registry, and the two matching test modules. Generated MARC1HTL-G1 accepts absent or one case-insensitive identity encoding, refuses all actual codings and 31/31 registered mutations, replays the exact 12+12 target-free cohorts, and cannot import the consumed executor or touch its old root. All 21 generated gates passed in 0.248252 seconds at 52,117,504-byte reported peak RSS with 215,460 temporary output bytes and zero real, network, payload, signal, target, model, score, or claim operations. The temporary outputs were removed.

The wrapper is not yet live-eligible. Its exact implementation commit must be pushed and both required CI jobs must pass before the one metadata-only attempt may run. This is the same confound-control to neural-positive-control to held-out-language path, not a pivot. Payload acquisition remains closed.

Exact wrapper 68ade0d4f6a58c19dbaae954a608080bdc6f128a then passed Base Python job 94089099869 and Optional Neuro Readers job 94089099850 in CI 31588920988 before the sole invocation. Read the consumed result and its machine registry. The standards repair worked: one 2,917-byte uncoded Wrist body with absent Content-Encoding passed transport and strict JSON parsing. Selection then parked at MARC1HTL-F04 because the live row count differed from the frozen 55-row contract. The actual count and rows were not retained.

MARC1-HT1A is consumed with no retry or rerun. It selected zero participants and opened zero payload bytes; all signal, target, model, prediction, score, and claim counters remain zero. Any metadata-snapshot repair is a separately named prospective lane and any new public body is a new Tier C event.

Result 1337a91ca2dd1f988ddcfc36631b7a1a8d832b0f passed Base Python job 94091696454 and Optional Neuro Readers job 94091696340 in CI 31589739739 before the zero-access MARC1-PG1 research. Pinned official Figshare OpenAPI source identifies the version-files endpoint as paginated, with a default page size of 10 and maximum of 1,000. The consumed wrapper omitted pagination, so partial first-page delivery is the leading hypothesis for MARC1HTL-F04; it is not proven because the actual row count was deliberately not retained.

The smallest prospective repair fixes the request identity to page=1&page_size=1000, requires the unchanged exact 55-row semantic inventory, and refuses a second page, fallback, partial cohort, or updated expectation. The next gate is to commit, push, and green this Tier A record, then freeze and qualify a generated-only pagination harness. No new live body, payload, signal, target, model, score, or claim action is authorized. This remains the same path toward cue-resistant neural evidence and later held-out language decoding, not a pivot.

Research 7a7883abda094eb9f202215b8b138a17cdff022e passed Base Python job 94095736694 and Optional Neuro Readers job 94095736770 in CI 31591022429 before the generated pagination contract was frozen. It permits only a future zero-network plan/qualify/inspect module after this exact contract is remotely green. Four accepted replay cases must share semantic and selection hashes; all 41 mutations must refuse across eight routes; all 18 acceptance gates must pass.

The contract retains exact page=1&page_size=1000 serialization and the unchanged 55-row identity. It has no execute command, URL or local-input argument, private-root name, automatic pagination, fallback, payload, signal, target, model, or scoring surface. The next gate is contract commit, push, and both green CI jobs. Generated success would still add no scientific evidence.

Contract ccb3ba8a839b3e6fc6844ad867ab0d5d295e20fb passed Base Python job 94098410925 and Optional Neuro Readers job 94098410868 in CI 31591853349 before the additive MARC1-PG1 implementation was built. The standard-library module exposes only plan, qualify, and inspect. Development MARC1PG-G1 passed 4/4 accepted cases, 41/41 refusals, and 18/18 gates while preserving the exact 55-row semantic identity and target-free 12+12 selection. It processed 1,019,776 generated input bytes in 0.089257 seconds at 40,091,648-byte reported peak RSS and removed all 183,355 temporary output bytes.

Exact implementation 2c98a2ad4b3972de5c2a398b85c0cf8735db89d4 then passed Base Python job 94104455930 and Optional Neuro Readers job 94104455857 in CI 31593790492 before the one registered generated closeout. It consumed at MARC1PG-F07 because its requested output parent /tmp is a symlink on macOS. The implementation had already loaded the contract, built both generated inventories, completed four accepted cases and selections, and constructed the generated private manifest in memory. No retry or corrected-path invocation is available.

The refusal created zero files and used zero real/private or network bytes. The next prospective gate is a separately named generated recovery that moves output-path preflight ahead of every contract and fixture operation. Only its own frozen contract, green implementation, and registered result could precede a later all-false Tier C packet. Payload, signal, target, model, and score work remains closed. MARC1-PG1 is consumed on the same cue- resistance to neural-positive-control to held-out-language path, not a pivot, and adds no scientific evidence.

Consumed result a4dcaea784f4c3a62547fd4f73bb3e2a5528100a passed Base Python job 94107907276 and Optional Neuro Readers job 94107907246 in CI 31594881048 before the artifact-only MARC1-OP1 recovery research. Its candidate design makes a held no-follow parent-directory capability the first operation, before repository read, contract load, pagination import, fixture construction, or selection. Device/inode binding and parent-relative creation keep later writes attached to the verified parent identity.

The local Python runtime exposes the required dir_fd, O_DIRECTORY, and O_NOFOLLOW primitives, but this is not a portability claim: missing support must refuse before generated work. The research performed one local runtime introspection and zero fixture, source, network, private, payload, signal, target, model, or score operations. Next: green this research before freezing the generated-only MARC1-OP1 contract. The scientific path is unchanged.

Research d02830b95c76bc428a297c6415db933452af5cbb passed Base Python job 94111539407 and Optional Neuro Readers job 94111539431 in CI 31595996923 before the generated recovery contract was frozen. It requires six accepted cases, 32 refusals across ten routes, and 20 gates. One future path-only probe must reach MARC1OP-P0 before one qualifier can seek MARC1OP-G1 at the exact /private/tmp destination.

The future module may defer-import hash-bound pure pagination helpers only after capability acquisition. It cannot call the consumed qualifier, modify its source, use absolute-path writes, contact a source, or expose payload, target, model, or score behavior. Next: green this exact contract before implementation. No scientific evidence changed.

Contract baade51 passed Base Python job 94115807028 and Optional Neuro Readers job 94115807008 in CI 31597291352 before the output-capability implementation. The additive standard-library wrapper acquires a held, no-follow parent capability before any repository or fixture work, then uses only relative exclusive writes and relative cleanup. Development MARC1OP-G1 passed all six accepted cases, 32 refusals, and 20 gates in 0.096 seconds at 33,767,424- byte reported RSS; all 184,173 temporary output bytes were removed. The registered path remains untouched until implementation CI is green. No live metadata, neural signal, target, model, score, or claim changed; this is the same route toward cue-resistant neural evidence and held-out language.

Exact implementation fcedcc3 passed Base Python job 94125013790 and Optional Neuro Readers job 94125013956 in CI 31600085119 before the registered result. The sole path-only preflight reached MARC1OP-P0 with zero experiment work, and the sole conditional qualifier reached MARC1OP-G1 using 184,173 temporary bytes in 0.097943 seconds at 33,882,112-byte reported RSS. Exact cleanup passed and both invocations are consumed. Next: green the result, then prepare only an all-false Tier C request for one bounded live metadata body. This generated result adds engineering trust, not neural evidence.

Result ca4679a passed Base Python job 94129199903 and Optional Neuro Readers job 94129199993 in CI 31601329375 before the MARC1-LM1 authorization packet was prepared. Every permission remains false. The packet proposes one additive generated/mock wrapper and, only after its exact implementation is separately green, one no-retry request to the immutable Figshare version-3 files endpoint with exact query page=1&page_size=1000. It allows one body capped at 2 MiB and a tiny target-free private inventory; it allows no EEG payload, target, model, score, or claim. Commit, push, and green this exact packet before identifying it as the sole Tier C gate. A fresh maintainer message is required after that identification. This remains the same thought-to-text research path.

After that identification, the maintainer supplied the exact words approved, continue, achieve a scientific claim, achieve thought to text 😎. Read the MARC1-LM1 decision. It binds only green request 4d3eb19, CI 31603530015, one future generated/mock wrapper, and one later metadata response with zero payload. Decision 060a365 passed Base Python job 94140250333 and Optional Neuro Readers job 94140250412 in CI 31604608307 before the MARC1-LM1 implementation.

The additive standard-library wrapper now acquires output authority first, validates the exact request and bounded transport, rejects target-like fields, binds the 55-row inventory and frozen 12-subject split, separates a private manifest from an aggregate report, and emits an aggregate failure receipt after a consumed marker. Final generated route MARC1LM-G1 passed four accepted transport forms, 36 refusals, and 20 gates using 184,466 generated input bytes and 19,030 temporary output bytes in 0.030280 seconds at 43,057,152-byte reported RSS. Exact cleanup passed. Real network, participant ZIP, payload, signal, target, model, prediction, score, and claim counters stayed zero.

First implementation push 8f67af2 failed both CI jobs because seven generated tests assumed macOS /private/tmp, absent on Linux. The corrected generated/test parent is the canonical real system temporary directory; the registered one-shot path is unchanged and was not touched.

Next: commit, push, and green this exact implementation. Only then execute the one registered metadata request; any post-marker failure consumes and parks, and success stops before payload. Thought-to-text remains the objective, not a predeclared outcome or permission to skip the controlled neural and held-out language gates.

Corrected implementation f9a1ece then passed Base Python job 94164152160 and Optional Neuro Readers job 94164152302 in CI 31611639130 before the sole registered metadata request. Read the consumed result. One 15,652-byte body passed bounded transport and strict JSON parsing, then the frozen inventory validator refused at MARC1LM-F04. The aggregate receipt does not identify the failed inventory predicate. No cohort, participant archive, EEG payload, target, model, prediction, or score was reached.

MARC1-LM1 is consumed with no retry or rerun. Next: specify a separately named prospective current-inventory identity lane before any new public response or payload decision. This is still the same path from trustworthy cohort identity to cue-resistant neural evidence to held-out language decoding and progressively stronger thought-to-text evidence.

Read the next same-path design, MARC1-SA1 source-aware inventory attestation. Official Figshare surfaces document a five-field public file core while also showing MD5 fields on some public metadata examples. The design therefore separates public core, optional provenance extensions, target-free cohort identity, and later acquired-byte SHA-256. Its 21 aggregate predicates and seven hashes can localize a future one-shot mismatch without publishing rows. No new dataset response or payload is authorized or accessed.

Research aa80503 passed Base Python job 94173234952 and Optional Neuro Readers job 94173234944 in CI 31614330447 before the MARC1-SA1 generated-only preregistration was frozen. It binds six semantic fixture families, 21 aggregate predicates, seven domain-separated hashes, 52 refusals, and 25 acceptance gates. The only future commands are plan, qualify, and inspect; there is no URL opener, execute mode, live path, payload, target, model, or score surface. Next: commit, push, and green this exact contract before generated-only implementation. A new live response remains a separate Tier C decision.

Contract 8f64ccb passed Base Python job 94180673330 and Optional Neuro Readers job 94180673125 in CI 31616551270 before the MARC1-SA1 implementation. The generated attestor reached MARC1SA-G1 with all six family routes, 52/52 refusals, and 25/25 gates. It processed 732,811 generated bytes and created then removed 109,589 output bytes in 0.052419791 seconds at 27,426,816-byte peak RSS. Next: green this exact implementation before one registered generated closeout. Live metadata and payload remain closed.

Exact implementation feb3b83 then passed Base Python job 94188922905 and Optional Neuro Readers job 94188922771 in CI 31619037335 before the registered MARC1-SA1 generated closeout. Its one MARC1SA-G1 run passed all six family routes, 52 refusals, and 25 gates in 0.053358083 seconds at 27,885,568-byte reported peak RSS. It used 732,811 generated input bytes, created and removed 109,589 temporary bytes, and accessed zero real metadata, payload, signal, target, model, or score. The generated closeout is consumed. A new source-aware live wrapper and one response require a separately green all-false Tier C packet and decision.

That all-false proposal is now frozen in the MARC1-SA1A source-aware live-metadata packet. It binds one future additive standard-library wrapper and one exact page=1&page_size=1000 response capped at 2 MiB. Green R1/R2 semantics may retain the frozen target-free cohort; historical drift or unknown extensions route R3/R4 and block selection. The request currently authorizes nothing, including wrapper implementation, network access, output writes, archives, payloads, EEG, targets, models, or scores. Next: green the packet, identify its exact commit/CI/scope, and wait for fresh packet-bound maintainer words.

Request b077550 then passed Base Python job 94198174069 and Optional Neuro Readers job 94198173901 in CI 31621794066. The maintainer's fresh instruction is now quoted byte-for-byte in the MARC1-SA1A authorization decision. It authorizes only the packeted metadata sequence: first green this decision, then build and green the generated/mock wrapper, then make the one bounded response check. Selective payload acquisition, neural experiments, scoring, replication, and language decoding remain closed behind later contracts.

Decision ef9ab91 then passed Base Python job 94353799568 and Optional Neuro Readers job 94353799602 in CI 31670457497 before the MARC1-SA1A source-aware wrapper was implemented. Generated MARC1SAL-G1 passed six source-schema families, all three accepted framing forms, 31 refusals, and 20 gates. It used 84,422 generated response bytes, created and removed 24,064 transient output bytes, and ran in 0.009288083 seconds at 37,552,128-byte reported peak RSS. Real metadata requests, archives, payload bytes, signals, targets, models, predictions, scores, retries, and claim upgrades stayed zero.

Next: commit, push, and green this exact wrapper in both CI jobs. Only then may its one fixed, 2-MiB-capped metadata request run. R1/R2 may expose only the frozen target-free cohort for a separately authorized acquisition packet; R3/R4 or any failure consumes and blocks the lane. Do not open an archive or infer a neural or language result from metadata compatibility.

Exact wrapper 74aff21 passed Base Python job 94360721568 and Optional Neuro Readers job 94360722170 in CI 31672761644 before the one source-aware metadata result. The request completed at MARC1SAL-R2 in 0.696664542 seconds at 33,439,744- byte peak RSS. It retained 23,112 bytes but selected zero subjects and opened zero participant archives. Payload, signal, target, model, prediction, scoring, retry, rerun, and claim counters remained zero.

MARC1-SA1A is consumed. The CLI did not expose the private R3-versus-R4 source route, and the already-inspected aggregate report plus private manifest were not reopened. Do not infer the specific difference, amend the source expectation, retry, or continue to Wrist payload acquisition. Next: close this cohort lane and compare independent cue-resistant datasets or prospective designs under new gates.

2026-08-21 MARC2-VR16A Generated Repair

The consumed private structural diagnostic ended at MARC2VR15P-R15, which identifies only a run-token width incompatibility. It did not retain the token or establish a real cohort.

The new generated-only VR16A adapter follows BIDS numeric-index semantics: it accepts ASCII [0-9]+, strips leading zeroes for semantic grouping, and keeps source-exact names for companion and storage accounting. Generated route MARC2VR16A-G1 passed 24/24 width/order/replay paths and 50 direct refusals in 2.224372999975458 seconds at 34,717,696-byte peak RSS. Six distinct lexical sources produced the same frozen semantic selection digest.

This is engineering evidence, not real-source or neural evidence. Exact implementation 6f92b84 passed both jobs in CI 32459984049. Tier A may now prepare an all-false private-confirmation packet, but any real confirmation, cohort freeze, archive member, neural payload, FW2, or CIL1 step requires that separate frozen Tier C packet and a fresh decision.

The all-false next request is now drafted as MARC2-VR16P. It proposes a generated fixed-path wrapper followed, only after separate proof barriers and fresh packet-bound authorization, by one target-free 418,755-byte structural read. The packet itself performs no private operation and grants no authority.

Request 619469a795a7044a4bbb77cef8986e0a7744473f passed Base Python job 96709074413, Optional Neuro Readers job 96709074203, and CI 32461465238. Its proof closeout now binds the three unchanged request artifacts totaling 32,234 bytes and performs zero implementation or private operations. The closeout must itself be committed, pushed, and remotely green before VR16P becomes the sole Tier C gate for a fresh packet-bound decision.

The maintainer's fresh packet-bound message was exactly continue. The VR16P authorization decision now binds only the unchanged green two-stage packet and exact decision 79d9b49 passed Base Python job 96714374482, Optional Neuro Readers job 96714374314, and CI 32463248736 before Stage 1.

The generated-only fixed-path wrapper now passes 24/24 width/order/replay paths and 71 direct refusals at route MARC2VR16P-G1. It processed 10,606,656 generated input bytes in 1.8574632920208387 seconds at 34,783,232-byte peak RSS, retained zero output, and performed zero private, archive, neural, target, model, score, FW2, or CIL1 operation. The complete dependency-light suite passes 4,519 tests with 204 expected skips.

Immediate gate: commit, push, and green the exact Stage 1 implementation, then commit, push, and green a proof-only closeout without repeating qualification. The sole 418,755-byte target-free structural read remains closed until both barriers pass.

Stage 1 commit 55f23d60621949ea008cca1e3ade80a3127cfc70 then passed Base Python job 96717830579, Optional Neuro Readers job 96717830410, and CI 32464397821. The proof-only closeout now binds those exact bytes and adds no implementation scope. It must itself be committed, pushed, and green in both jobs before the one registered private invocation becomes available.

Proof closeout 865d76aff51842e3b57600a7dab399d2bbe91d2e passed both jobs in CI 32465104587 before the sole invocation. The command read and strict- parsed exactly 418,755 target-free structural bytes once, called VR16A once, and consumed at MARC2VR16P-R4 without freezing a cohort. R4 means only numeric identity, exact task token, or companion validation refused. The exact predicate, value, filename, path, row, participant, selection, and cohort are unavailable and must not be inferred.

VR16P has no retry, rerun, repair, fallback, substitution, or private reinspection. The next safe lane is artifact-only/generated-only R4 predicate decomposition. FW2 and CIL1 remain ineligible; no archive, neural payload, target, model, prediction, score, or scientific claim is open.

That next lane is now frozen as MARC2-VR17A. It binds 13 committed inputs totaling 204,240 bytes and tests whether the green VR15P width result, VR16A adapter semantics, and prior producer evidence reduce R4 to four generated task/companion classes. The reduction remains a hypothesis until all generated composition gates pass. Commit, push, and green the exact registration before implementation. The lane has no private executor and authorizes no new read.

Registration e1c9366 passed both jobs in CI 32467147580 before one bounded generated preflight. That preflight caught a preregistration error: the two already-supported controls pass both VR15A and VR16A, while only the four extended-width variants move from VR15A R15 to VR16A success. All six preserve one semantic selection digest.

VR17A is therefore parked at MARC2VR17A-P01; its full equivalence and residual matrices did not run. The measured preflight used 2,651,670 generated bytes, 0.6208636660012417 seconds, 33,390,592-byte peak RSS, and zero retained output or private operations. Next: preregister corrected MARC2-VR17B with the observed 2-control/4-repair mapping. No real cohort or neural evidence exists in this lane, and FW2/CIL1 remain closed.

Corrected MARC2-VR17B is now preregistered against green parked-result commit 48775bb / CI 32468221097. It freezes the observed two G1-to-G1 controls and four R15-to-G1 repairs before replaying the full generated equivalence and four-class residual matrices. Eight focused and 4,550 complete dependency- light tests pass with 204 expected skips.

Immediate gate: commit, push, and green this exact registration before any VR17B implementation. It has no private executor and opens no cohort, neural, target, model, score, FW2, CIL1, release, or scientific claim.

Registration cde8569 passed both jobs in CI 32469173279 before a five-case generated preflight. Two residual expectations were wrong: wrong-task is reported as F04 / core identity differs, and the old collision witness is caught first as F05 / companion run spelling differs. VR17B is parked at MARC2VR17B-P01; its full matrix did not run and no implementation module was added.

Next: freeze VR17C with the observed task reason and a true same-token, distinct-name normalized-collision witness. This remains generated structural work, with no private cohort or neural evidence and no FW2/CIL1 authority.

That successor is now frozen as MARC2-VR17C against green VR17B result 3d69d65 / CI 32470024380. It preserves the 1,227-row envelope and exact run token while creating a distinct _acq-copy companion, which reaches the intended normalized-collision branch in generated development. The registered full matrix remains 24 paired equivalence paths plus 20 residual paths.

Eight focused and 4,564 complete dependency-light tests pass with 204 expected skips. Immediate gate: commit, push, and green the registration before implementation. No private, neural, model, score, FW2, or CIL1 work is open.

Registration a34896d passed both jobs in CI 32470828824 before VR17C implementation. The generated-only classifier now passes 24 paired equivalence paths, 20 residual paths, all four hypotheses, and 50 direct refusals. G1 and R1-R4 each appear four times under both orders and two exact replays.

The measured pass used 19,213,944 generated bytes in 3.7416582500445656 seconds at 36,569,088-byte peak RSS, emitted 2,719 aggregate bytes, and retained zero. Twenty focused and all 4,576 dependency-light tests pass with 204 expected skips. Immediate gate: commit, push, and green this exact implementation/result, then add a proof-only closeout. No private source, cohort, neural payload, target, model, score, FW2, or CIL1 work is open.

Exact VR17C implementation fcff514 passed both required jobs in CI 32472572881. Its proof-only closeout now binds the exact committed module, tests, measured documents, and preproof registries while repeating no explicit qualification and touching no private state. Commit, push, and green this closeout before preparing any future private-discriminator packet.

Proof closeout 820592e then passed both jobs in CI 32473172345. The next all-false request is prepared as MARC2-VR18P: generated wrapper qualification first, then only after separate proof barriers and fresh Tier C authority one 418,755-byte target-free structural read. Every request flag is false now. Commit, push, and green the request and its later proof closeout before any implementation or private access.

Exact VR18P request 521f1de passed both required jobs in CI 32474183647. Its proof-only closeout now binds the three unchanged request artifacts totaling 32,886 bytes and records zero implementation, private, or scientific operations. Commit, push, and green this closeout before naming VR18P as the sole Tier C gate. A fresh maintainer decision is still required after that proof; no earlier continuation is retroactive.

Proof closeout ea5b4c7 then passed both jobs in CI 32474864890. After VR18P was named as the sole gate, the maintainer's next exact message was continue. The packet-bound decision records those eight bytes and binds only the unchanged two-stage request. Commit, push, and green the decision before Stage 1 generated wrapper work. No readiness or private path is open yet.

Decision 5113be7 passed both jobs in CI 32475765286 before Stage 1. The generated fixed-path wrapper passed 20 matrix paths, one temporary success rehearsal, and 82 refusal checks. G1 and R4-R7 each appeared exactly four times. The pass used 9,037,650 generated bytes in 1.4131330830277875 seconds at 37,879,808-byte peak RSS and retained zero.

All 4,620 dependency-light tests pass with 204 expected skips. Immediate gate: commit, push, and green the exact implementation/result, then green a separate proof-only closeout. The implementation proof is null now, so execute refuses before readiness or private access. No neural or scientific result was created.

Exact VR18P Stage 1 6688123 passed both required jobs in CI 32477528982. The proof-only closeout now binds that exact implementation and seven preproof Git blobs without rerunning qualification or touching private state. The production wrapper and measured generated result are unchanged; only the proof-null refusal test and proof record test advance with the registry.

Next gate: commit, push, and green this closeout. Only then may the one already decision-bound target-free structural invocation run. Until that proof, readiness, .codex_work, the private source, and the real output root remain closed. FW2/CIL1 and every neural or scientific claim remain closed.

Proof closeout 535bd9b passed both required jobs in CI 32478506443 before the sole registered execution. That invocation consumed at MARC2VR18P-R4 after one 418,755-byte target-free structural read, one strict parse, one VR16A call, and one frozen-map lookup. Runtime was 10.053867666982114 seconds, peak RSS 27,443,200 bytes, output 2,063 bytes, and network/new payload bytes zero.

R4 narrows the blocker to the core task-or-identity class but does not say which of those failed and does not identify any private value, row, path, person, selection, or cohort. No cohort was frozen, so FW2/CIL1 remain ineligible. VR18P has no retry or reinspection. Next: freeze an artifact-only, generated-only decomposition of that class before considering any new Tier C private packet.

That decomposition is now frozen as MARC2-VR19A. Static source analysis finds two F04 producers in the exact validator, both guarded by task != "freewill"; the consumed R4 pair selects the earlier translated reason. Eight generated cases will test that implication across two orders and two replays after registration CI is green. All 4,637 dependency-light tests pass with 204 expected skips.

Immediate gate: commit, push, and green the exact VR19A registration. It is artifact-only and generated-only. No private token, cohort, neural payload, model, score, FW2/CIL1 execution, or scientific claim is open.

Registration 9365b0f passed both jobs in CI 32480420157 before implementation. VR19A then passed all 32 generated paths, all four hypotheses, and 40 direct refusals. G1/R1/R2 counts were exactly 4/16/12 across two orders and two replays.

This tightens the real consumed structural result: under exact VR16A code, R4 means the task token was not exactly lowercase freewill, not a generic identity failure. The token itself remains unavailable. The pass used 13,748,732 generated bytes, 1.759938333008904 seconds, 51,429,376-byte peak RSS, 2,326 output bytes, and zero retention.

Immediate gate: commit, push, and green the exact implementation/result, then green a proof-only closeout. No private read, cohort, neural evidence, model, score, FW2/CIL1 execution, or scientific claim is open.

Exact implementation fda3a3a passed both required jobs in CI 32481785128. The proof-only closeout now binds that commit, both job IDs, the two preproof registry hashes, and seven exact Git blobs. It repeats no qualification and performs zero private operations.

Immediate gate: commit, push, and green this exact closeout. After that, VR19A is closed as generated structural engineering. Any private task-token read remains a new Tier C packet; no cohort, neural evidence, model, score, FW2/CIL1 execution, or scientific claim is open.

Closeout 801fce3 passed both jobs in CI 32482429485. Primary-source review then found the blocker: the dataset paper publishes raw BIDS filenames as task-reachingandgrasping_run-zzzz, but the inherited MARC2 selector requires task-freewill. That public mismatch explains why the one-shot private source reached the non-freewill structural class without exposing a private row.

Immediate gate: commit, push, and green the Tier A task-identity research. Then preregister an additive generated-only source-exact selector repair. No private source, archive member, neural payload, model, score, FW2/CIL1 execution, or scientific claim is open.

Research d5556d6 passed both jobs in CI 32483267516. The generated-only repair is now frozen as MARC2-VR20A: exact lowercase task-reachingandgrasping, source-exact filenames and run spellings, 20 success paths, at least 50 direct refusals, one CPU thread, and zero retained output.

Immediate gate: commit, push, and green the VR20A registration before implementation. A private confirmation packet is not eligible until the exact implementation/result and proof-only closeout are remotely green. Keep .codex_work, VR18P consumed state, and the unrelated tracker inspection NDJSON untouched.

Registration cd71807 passed both jobs in CI 32484725113. Generated route MARC2VR20A-G1 now passes 20 source-exact paths and 53 refusals for exact task-reachingandgrasping, selecting the same 16 subjects / 96 bundles / 384 members across every valid run spelling. Runtime was 1.7907995419809595 seconds at 35,143,680-byte peak RSS with zero retained output.

Immediate gate: commit, push, and green the exact VR20A implementation/result, then add a proof-only closeout without rerunning qualification. Keep .codex_work, VR18P consumed state, and the unrelated tracker inspection NDJSON untouched. No private cohort, neural payload, or scientific claim is open.

Exact implementation bf4d2b7 passed both jobs in CI 32486620566. The proof-only closeout now binds that commit, its preproof registry hashes, and seven Git blobs without repeating qualification or touching any private path.

Immediate gate: commit, push, and green the exact proof-only closeout. Only after that remote proof may Tier A prepare a separate all-false private structural-confirmation packet. That packet would still require a fresh Tier C decision before any private read. Keep .codex_work, VR18P consumed state, and the unrelated tracker inspection NDJSON untouched.

Proof-only closeout 9b5bea4 passed both required jobs in CI 32487854026. The next gate is now prepared as MARC2-VR20P. Read docs/MARC_2_PUBLISHED_TASK_PRIVATE_CONFIRMATION_AUTHORIZATION_PACKET.md and registries/marc2_published_task_private_confirmation_authorization_request.v0.json. It is an all-false two-stage request: generated fixed-path wrapper qualification first, then only after every proof barrier and a fresh Tier C decision one 418,755-byte target-free structural read and one exact VR20A call.

Commit, push, and green the exact request, then create and green a non-scope- changing request proof closeout. Only afterward identify VR20P as the sole active Tier C gate. Before a fresh decision, do not implement the wrapper or touch readiness, .codex_work, the private source, or an output root. FW2, CIL1, archive payload, neural data, targets, models, predictions, scores, release, and scientific claims remain closed.

Exact request bef2391d8edf92c5edf8a3624831e50430636626 passed Base Python job 96793861959, Optional Neuro Readers job 96793861717, and CI 32489589922. Read docs/MARC_2_PUBLISHED_TASK_PRIVATE_CONFIRMATION_REQUEST_PROOF_CLOSEOUT.md and its machine proof. The closeout binds the unchanged 35,823 request bytes and performs zero implementation, readiness, private, or scientific operation.

Commit, push, and green this exact proof-only closeout. Only afterward may VR20P be identified as the sole active Tier C gate. A fresh packet-bound maintainer decision and its own remote proof are still required before Stage

Do not touch .codex_work, the private source, consumed VR18P state, or a real output root.

COMM-G1 Prospective Control Correction

The COMM-G1 experiment registration is remotely green. Before implementing it, read docs/COMM_G1_GENERATED_EXPERIMENT_AMENDMENT_1.md: the amendment replaces a class-preserving EEG derangement with a source-only cyclic class rotation that has no fixed points. This correction occurred before any COMM-G1 implementation, qualification, prediction, target delivery, or score.

The generated implementation is now recorded in docs/COMM_G1_GENERATED_IMPLEMENTATION.md and registries/comm_g1_generated_implementation.v0.json. Its disposable local measurement stayed under every cap and exercised 35 refusals, but it is not the official generated qualification and has no scientific value. Exact implementation 5621fb8e378def319f6529e3d3173ffc88dbe9af is remotely green.

The sole official qualification subsequently executed and is consumed. Read docs/COMM_G1_GENERATED_QUALIFICATION_CLOSEOUT.md and registries/comm_g1_generated_qualification_closeout.v0.json. The binding route is COMM-G1-R0 because the replay digest was incomplete, the two replays were not isolated in clean workdirs, and two named refusal families were not explicitly exercised. Do not rerun or repair COMM-G1. Its score is not accepted and it has no scientific value.

Next read docs/COMM_G2_GENERATED_PROOF_QUALIFICATION_PREREGISTRATION.md, registries/comm_g2_generated_proof_qualification_contract.v0.json, and the binding docs/COMM_G2_GENERATED_PROOF_QUALIFICATION_CLOSEOUT.md. COMM-G2 was a new lane, not a COMM-G1 repair. It froze the exact scientific module and added only stronger proof isolation, complete digests, target separation, filesystem publication guards, and full named refusal coverage.

The registration is green at 38f5e725f46ad7a8a18ff92dae665a1569e13d07. The separate implementation is documented in docs/COMM_G2_GENERATED_PROOF_IMPLEMENTATION.md and registries/comm_g2_generated_proof_implementation.v0.json. Exact implementation 83658976fb334a6b125a441af5dbc8cccf416e1f passed CI 33056826167 before the one official invocation. That invocation refused with G2-CHILD-TIMEOUT before the first replay returned, published no official result, and left no temporary directory. COMM-G2 is consumed at COMM-G2-R0; do not rerun, repair, overwrite, or reinterpret it as accepted.

FilesExpand file tree

START_HERE.md

Latest commit

History