Status: work order 4 complete; one measured symbolic roundtrip consumed
Date: 2026-08-08
Result registry:
registries/classical_eeg_adapter_result.v0.json
The three-adapter contract was committed at ea5fafda3e408972797e579f00d786ab6c8ee6bc
and passed push CI 31279856066. The implementation was committed at
eefb7b066810c2a6b87417b105bdb746218e87dc and passed exact-SHA push CI
31280581308 before the measured roundtrip began. Base Python passed in 16
seconds and Optional Neuro Readers passed in 57 seconds.
Exactly one symbolic plan creation and one inspection then ran in an automatically removed temporary directory. No retry or post-result tuning occurred. The plan was hashed and measured before removal; it is not retained in Git or elsewhere in the workspace.
| Field | Result | Cap |
|---|---|---|
| Runtime | 0.12 seconds | 15 seconds |
| Peak RSS | 22,822,912 bytes | 268,435,456 bytes |
| Input contract | 12,025 bytes | recorded |
| Generated plan | 27,335 bytes | 1,048,576 bytes |
| Retained generated plans | 0 | 0 in Git |
The serialized plan SHA-256 was
680f95e41ea2e4b7b2bcec0961d493c9bc2889294f0652dabc120db79f6ecb2d.
Its canonical content hash, excluding only the self-referential hash field,
was 66800348e76d03b9b994a460b2e78fbe569c450fdb289be5948cecbcea860bf1.
The plan contained 96 target-free symbolic items in 48 pair-bound groups: 48 train items in 24 groups, 32 check items in 16 groups, and 16 final items in 8 groups. Creation and inspection returned zero and emitted equal compact summaries. No adapter winner was selected.
The registered families remain low-frequency shrinkage LDA, causal 8-30 Hz CSP-LDA, and regularized Riemannian MDM. Their fit stages are restricted to train groups or data-independent configuration. Check and final targets remain unavailable to fit or transformation. The plan requires zero right-context and zero post-event samples, but no causal producer was executed and end-to-end latency was not measured.
One symbolic plan build and one symbolic inspection occurred. Adapter backend imports, optional dependency installations, array reads, raw or protected data reads, public EEG reads, target or label-value reads, feature extraction, parameter updates, inference, training, scoring or selection, network or provider calls, stream/device/hardware operations, and scientific-claim upgrades were all zero.
The deterministic replay, exact hash binding, pair/group isolation, train-only fit firewall, all twelve malformed-plan refusals, bounded read, exclusive create, collision refusal, standard-library-only import, and CLI equality gates passed. Closeout verification passes 39 focused checks and the complete suite at 1,279 tests, 3 expected skips, and 469 subtests. Ruff, compileall, all registry JSON, root and command help, and diff hygiene also pass.
Warnings remain explicit:
- this is a symbolic interface plan, not EEG feature extraction;
- no classical adapter was selected or executed;
- synthetic identities are leakage-test surrogates, not participant evidence;
- causal fields constrain future work but do not measure a producer or latency; and
- public or protected adapter execution remains a separately gated Tier C action.
All 18 acceptance gates passed. Work order 4 is complete and consumed without retaining generated output. Work order 5 may now freeze and synthetically test contact-mask, channel-noise, and missing-channel semantics for an ear-channel adapter. That route authorizes no hardware, real data, model fit, inference, or scientific scoring.
Engineering capability added: NeuroDecodeKit can deterministically construct, hash, save, inspect, and fail closed on leakage in three optional classical EEG adapter plans.
Scientific claim not established: no EEG adapter was executed and no real neural effect, decoding accuracy, generalization, latency, device performance, home use, or clinical result was established.