Repository navigation
Description
Current state: Proposed. This RFC defines a standard result bundle and report contract for RSI tasks; implementation is pending review.
Decision needed: Repository maintainers should decide whether to adopt the four delivery files, result semantics, and rollout below, and designate the individual decision and acceptance owner.
Next action: Maintainers record adoption, requested changes, or deferral and assign owners to any accepted implementation tasks.
Scope: InferRSI project optimization, Benchmark optimization sessions, and exports of existing experiment records.
Proposal
Each RSI task should deliver a result bundle that people can read directly, programs can consume, and reviewers can trace to source code and validation evidence. Standardize four entry points:
| File | Audience | Purpose |
|---|---|---|
report.md |
Users and reviewers | Default reading entry point: goal, outcome, optimization history, delivered code, and limitations |
report.html |
Users who need charts and interaction | An offline report generated from the same result snapshot as Markdown |
campaign-result.json |
Automation and subsequent agents | A versioned result contract for the complete experiment |
artifacts.json |
Reviewers and archive tools | An index of code, raw measurements, validators, and execution records |
The final response should lead with the conclusion and report.md, followed by the HTML report, structured results, and code location. Label HTML links “Open in a browser” so users who follow a file link in an IDE know how to view the rendered report.
Generate these files from the same retained records. Existing modules continue to own experiment control state, assessments, and raw evidence. Report export must not create a second optimization state, rerun measurements, or change acceptance gates.
Background and current behavior
Existing entry points retain much of the required information, but delivery depends on the execution path:
| Existing path | Existing artifacts | What this proposal adds |
|---|---|---|
Iterative campaign and benchmarks session optimization |
plan.json, state.json, summary.json, and candidates and evidence under rounds/ and final/ |
A stable experiment-level delivery contract and a default readable report |
dashboard --input |
Offline HTML generated from existing records | A delivery entry point bound to the same snapshot as Markdown and JSON |
Legacy run and report |
JSON, logs, and report.html for one A/B experiment |
Adapter-based export that preserves the scope of a single experiment |
| Apps operation output | Existing result.json, manifest.json, and fixed-version references |
References to these results without treating an operation outcome as the conclusion of an entire optimization task |
The iterative workflow's summary.json already contains the final assessment, selected candidate, per-round curves, elapsed time, and imported usage. It is a summary rebuilt from control records. Apps result.json belongs to the operation-level contract. The new experiment-level file is therefore named campaign-result.json; it does not replace either existing file.
This RFC defines delivery and presentation requirements. New optimization algorithms, statistical tests, and repository-wide performance thresholds are outside its scope. PDF, Word, online hosting, and automatic publication to external systems are excluded from the initial implementation.
The related artifact archiving RFC addresses repository inclusion and external retention. This proposal defines what users receive and how results are represented; it does not require large delivery artifacts to be committed to Git.
Bundle layout and generation points
Use the following export layout. This layout does not require the runtime to rearrange existing evidence:
delivery/<export-id>/
report.md
report.html
campaign-result.json
artifacts.json
attachments/ # Optional patches, small configurations, and evidence
The four entry files form the minimum result bundle. Code and raw evidence may be included or delivered through fixed-version references. Reference-based delivery must explain access requirements and missing items. A report with external references can be read offline without necessarily supporting complete offline verification.
Generation follows these rules:
- After an iterative optimization completes its independent
finalevaluation, generate the final report automatically, including when the verdict is FAIL or no improvement was found. - Failure, interruption, budget exhaustion, or an unavailable final evaluation must still allow partial export. The first page must state that the task is incomplete and explain why.
- Allow on-demand progress snapshots during execution. Record their observation cutoff and distinguish them from final results.
- Export reads retained results only. It must not launch a missing final evaluation merely to produce a report.
- After a forced process termination, allow a separate export of retained records. Automatic report creation cannot be guaranteed for every termination path.
Each export has a distinct export_id and identifies its source record revisions or digests. For an active task, read a consistent committed snapshot or detect changes and retry the read. Do not combine a plan, candidate, and final evaluation from different snapshots.
Experiment-level result contract
campaign-result.json uses UTF-8 JSON. The first version has schema_id: inferrsi.campaign-delivery and schema_version: 1. Every top-level field below must be present. Unavailable data is represented by null, with the reason recorded in gaps. An empty array means the collection is known to be empty; null means it is unknown.
| Field | Type | Required meaning |
|---|---|---|
schema_id, schema_version |
string, integer | Delivery contract identity and version, independent of source data versions |
export |
object | id, generated_at, snapshot_at, kind, and generator version; kind is progress or final |
campaign |
object | Experiment ID, goal, task kind, acceptance basis, and frozen-plan reference; distinguish project optimization, Benchmark optimization, a single experiment, and external records |
sources |
array | Fixed source record references resolved through artifacts.json, retaining source types and record revisions |
execution |
object | Original execution state, budget, completed rounds, planned rounds or round limit, round mode, stop reason, and start/end times |
completion |
object | status: completed / incomplete and a reason; describes completion of required workflow steps only |
outcome |
object | kind: improved / baseline_retained / no_valid_candidate / undetermined and the selection rationale |
baseline |
object or null | Baseline source identity, effective configuration, and evidence references |
selected_candidate |
object or null | Retained selected code, configuration, fingerprint, and patch or retrievable revision reference |
final |
object or null | Independent reevaluation of the selected code, its fingerprint, protocol reference, and completion facts; null when not run |
acceptance |
object | Primary claim, scope, assessment reference, and original verdict; distinguish assessed and unassessed results as specified below |
metrics |
array or null | Interpretable, traceable comparisons separated by workload, concurrency, phase, and statistical definition |
checks |
array or null | Execution state, verdict, basis, and evidence references for quality, correctness, performance gates, and other checks |
rounds |
array or null | Ordered index of submitted attempts, including successful, failed, interrupted, and rejected candidates |
usage |
object | Wall time, observed tokens, and available device usage and cost; each with its own source and coverage |
gaps |
array | Incomplete steps, unverified scope, unknown provenance, and missing or corrupt evidence, with their effects |
artifacts_ref |
string | Relative path to this bundle's artifacts.json |
execution preserves the source state and its namespace. An existing blocked state must not be rewritten as a candidate error. completion is a projection of workflow completion for delivery, not a state written back to Harness. A final snapshot means the task represented by that delivery has stopped; it does not imply that all required steps completed.
All timestamps use ISO 8601 strings with a time zone. Numeric fields contain finite numbers or null. Do not use 0, an empty string, NaN, or Infinity to represent an unobserved value. JSON retains calculation precision; reports may round for display.
Each rounds entry includes at least the round number, attempt ID, candidate fingerprint, hypothesis, change summary, original execution state, assessment reference, metric references, selection decision, and best candidate at that point. Retries have distinct attempt IDs and retain their associated round; retry counts must not be treated as completed-round counts. Each checks entry retains the check name, whether it is required, execution state, frozen-gate reference, observation, verdict, and evidence references. An unexecuted check cannot be shown as passed.
Completion and acceptance semantics
Answer three questions separately: whether the workflow completed, which code was selected, and which acceptance conditions passed.
| Situation | Completion and selection | User-facing result |
|---|---|---|
| Required rounds or a permitted early stop are complete, the selected candidate completes independent final evaluation, and primary acceptance passes | completed; improved when the improvement conditions are met |
Original PASS, its scope, and final metrics |
| Assessment completes under the frozen protocol but demonstrates a quality or performance-gate violation | May be completed if all workflow obligations are fulfilled |
Original FAIL; completing execution does not establish optimization success |
| Search finds no valid improvement and the baseline completes the required independent final evaluation | completed, baseline_retained |
“Baseline retained; no verified improvement,” together with the actual acceptance verdict |
| Final evaluation is missing or times out, evidence was incomplete at assessment time, or execution facts cannot be established | incomplete; retain a supported selection outcome or use undetermined |
Incomplete acceptance or the original INCONCLUSIVE, with the gap explained |
| Search produced a PASS, but the selected version fails final evaluation | Preserve search records; derive completion from fulfilled workflow obligations | Final FAIL or INCONCLUSIVE; do not substitute the search PASS |
acceptance.status is assessed or not_assessed. For an assessed result, preserve the assessor's PASS / FAIL / INCONCLUSIVE verdict. For an unassessed result, verdict is null and the report says “Not assessed.” Do not fabricate an INCONCLUSIVE assessment record. When the source contains an actual INCONCLUSIVE assessment, display it normally.
For outcome.kind, improved requires the selected non-baseline version to pass its required final evaluation and meet the frozen protocol's improvement conditions. baseline_retained records the decision to keep the baseline; acceptance is reported separately. no_valid_candidate means existing assessments explicitly rule out all deliverable versions. Use undetermined when the evidence cannot establish one of those outcomes. The best search candidate may still be delivered through selected_candidate, but its lack of final qualification must be explicit.
Bind acceptance to claim_id, scope, and assessment_ref. If historical data lacks these identities, preserve its original verdict, record identity_incomplete, and do not invent IDs. List multiple claims separately. The frozen plan determines the primary claim; the exporter must not select whichever claim has the most favorable result.
Acceptance conditions come from the frozen protocol. Missing a required independent final evaluation or validation level prevents a claim that the complete iterative optimization passed. Single experiments and external results retain their original verdict and scope; they do not automatically become final acceptance of an iterative task. L0–L2 results must not be generalized into end-to-end service gains.
If evidence for a historical PASS later becomes unavailable, preserve the historical completion state and PASS while prominently reporting the current verification gap. Do not silently rewrite history or present unavailable evidence as accessible.
Binding delivered code to final evaluation
selected_candidate identifies a retained snapshot, not a working directory that may continue changing. Provide either a retrievable exact revision and configuration, or a baseline revision plus a patch and content fingerprint. Repository HEAD alone is insufficient when uncommitted changes contributed to the result.
final binds the same candidate fingerprint and frozen protocol. If fingerprints differ, record the binding failure in gaps and do not claim final acceptance for that selected code. Preserve the original assessment as an audit fact.
Delivering the baseline is valid when no improvement was found. selected_candidate: null is valid when no code has been selected. Generated audio, images, or video are quality or outcome evidence and belong behind artifact references rather than inline in the summary JSON.
Metrics and usage
Each metric includes at least name, unit, direction, scope, workload_id, configuration or concurrency identity, statistic, baseline value, candidate value, sample and repetition counts, measurement phase, evidence references, and comparability. Distinguish search and final phases.
Calculate the following display values only when baseline and candidate protocols, workloads, statistical definitions, and units are comparable, and both values are finite and positive:
| Direction | speedup |
improvement_pct |
|---|---|---|
| Lower is better | baseline / candidate | (baseline − candidate) / baseline × 100 |
| Higher is better | candidate / baseline | (candidate − baseline) / baseline × 100 |
Both values identify their baseline and measurement scope. If values are missing or incomparable, leave derived gains null and explain why. Preserve raw values and task-specific assessments for zero-valued metrics or metrics where a ratio is inappropriate. Do not produce an aggregate speedup across workloads without a declared aggregation method.
Optimization curves retain each round's valid result and best-so-far value. Failed rounds show their failure reason and must not be filled with a zero score. Only an independently reevaluated selected result can establish final gains. When the baseline is retained, report zero optimization gain and an optimization speedup of 1.0 only after the required final evaluation completes and supports that statement. Preserve the independently measured A/A ratio separately.
Each usage measure records value, source, and coverage: complete / partial / unknown. If only some rounds have token observations, label the observed subtotal with its coverage. Missing usage is not zero. Do not add whole-session usage to per-round usage for the same work, and treat cached input as a subset of input tokens. Device time and monetary cost require actual observations or an explicit accounting basis; do not infer them from wall time.
Example fragment
This fragment illustrates the status semantics of an incomplete task. It is not a complete result file or actual experiment data:
{
"schema_id": "inferrsi.campaign-delivery",
"schema_version": 1,
"completion": {
"status": "incomplete",
"reason": "Search stopped before independent reevaluation of the selected version completed"
},
"outcome": {
"kind": "undetermined",
"reason": "Search results do not yet establish a final improvement"
},
"final": null,
"acceptance": {
"status": "not_assessed",
"verdict": null,
"claim_id": null,
"scope": null,
"assessment_ref": null
},
"gaps": [
{
"code": "FINAL_NOT_RUN",
"affects": "final_acceptance",
"reason": "No independent final evaluation record is available"
}
],
"artifacts_ref": "artifacts.json"
}Report reading order
Markdown and HTML use the same core sections, presented in this order:
- Result summary: Optimization goal, workflow completion, final acceptance, selected code, gains, and current evidence gaps.
- Experiment conditions: Model, hardware, runtime, baseline and candidate revisions, workload, concurrency, quality constraints, and frozen protocol.
- Final comparison: Comparable baseline and candidate values, gains, sample counts, repetitions, and available variability information. Do not claim statistical significance without supporting analysis.
- Optimization history: Each round's hypothesis, actual change, result, and adoption or rejection rationale, including unproductive, regressing, and failed directions.
- Validation and limitations: Quality checks, final evaluation, unverified scenarios, applicable scope, and raw evidence locations.
- Code and reproduction: Selected revision, necessary configuration or patches, measurement and validation entry points, and external model and environment dependencies.
- Time and resources: Actual elapsed time, stop reason, observed usage, and coverage.
The first screen should present conclusions rather than only raw JSON or a log index. Core conclusions and tables in report.md must be readable in an ordinary Markdown viewer. HTML embeds its presentation data and styles; core conclusions remain readable without scripts, while interactive details may degrade. State separately whether original evidence is included in the bundle.
Render conclusions, numbers, units, and validation states from the structured result. Agents may add explanatory analysis, but must label it as analysis and cite its evidence. Narrative editing cannot promote an acceptance verdict. Both reports display export_id, the snapshot time, and the structured result's digest so readers can confirm that they belong to the same export.
Evidence index and consistency
artifacts.json has its own version identity and records export_id. Each entry includes a stable ID, role, location, media type, size, SHA-256, source record revision, and availability. Roles cover plans, protocols, source code, patches, configuration, measurements, quality checks, final evaluation, logs, profiles, generated media, knowledge references, and report files.
Use relative paths for bundled files. External artifacts use fixed-version references and content digests. Mark access requirements for restricted artifacts without placing credentials in references. Unknown locations may be null with an explicit gap. Retain known digests when an artifact is temporarily inaccessible.
availability distinguishes available / missing / restricted / not_checked; integrity distinguishes verified / mismatch / not_checked. Record when checks were performed. Metadata or reference-format validation alone does not establish verified bytes. Matching a file hash establishes content identity, not experimental correctness.
Generation order is: fix the source snapshot → generate campaign-result.json → render both reports → generate the artifact index → validate and publish. The index includes digests for the first three files but does not hash itself. campaign-result.json references the index by path only, avoiding a digest cycle.
A failed export must not replace the previous complete bundle. Publish to a new destination directory and confirm success. Atomic visibility and power-loss durability retain the platform limits of the existing Apps publication facilities. On publication failure, preserve committed execution and assessment references; a report failure must not trigger experiment reexecution.
Large collections of per-request data, logs, profiles, and generated media may be archived externally. Export and storage migration must not delete failed samples, modify frozen validators, or overwrite historical assessments. Record redaction scope and original-artifact access restrictions separately for shared copies.
Compatibility and implementation
The initial implementation adds result projection and report generation in Apps using existing record readers. Harness continues to own execution and selection state, assessors continue to supply acceptance decisions, and Apps assembles and presents the delivery bundle.
Proposed export entry point, not an existing implemented command:
inferrsi export --input <campaign-or-run-directory> --output <new-delivery-directory>This command exports retained records only. Preserve compatibility with existing dashboard, legacy report, Apps operation contracts, and their exit codes. The new export command's exit code describes export success or failure; consumers read the experiment verdict from the result. Successfully exporting a FAIL report is not an export failure.
Integrate the same exporter into the iterative final workflow, returning execution and export outcomes separately. A report-generation failure must not roll back a committed final evaluation. Users can resolve the output problem and export again.
Use separate adapters for historical summary.json, Apps result.json, and external records. Preserve original files and explicitly record source fields and mapping gaps. Legacy single A/B experiments must not acquire invented rounds or independent final evaluations. Externally reported results must not be presented as locally revalidated.
Delivery JSON is a rebuildable, read-only projection. recover and selection logic must not consume it as a new control state. The same source snapshot and generator version should yield identical metrics and conclusions; export timestamps and IDs may differ.
Contract compatibility rules: adding optional fields retains the version; renaming fields or changing types or semantics requires a version increment. Readers must validate schema_id and explicitly reject unsupported versions rather than infer meaning from similarly named fields.
Acceptance criteria
Implementation acceptance should cover these observable behaviors:
| Scenario | Expected behavior |
|---|---|
| Normal completion with final acceptance | All four entry files exist; code, final evaluation, protocol, and verdict correspond; metrics agree across the three result views |
| Assessment completes without meeting the goal | Deliver the actual FAIL or baseline-retained outcome; execution completion does not imply optimization success |
| Search records only, or final evaluation fails or times out | Show workflow gaps or the final verdict; a search PASS cannot substitute for final acceptance |
| Historical single experiment or external record | Export preserves the original scope without inventing final evaluation, model revisions, or validation levels |
| Missing usage or observations covering only some rounds | Values and coverage remain accurate; unknowns are not zero; session and per-round usage are not double-counted |
| Incomparable measurements or an inapplicable ratio | Preserve raw values, leave derived gains null, and explain why |
| Selected candidate and final fingerprints differ | Show the binding failure; do not attribute the assessment to different code |
| Evidence for a historical PASS becomes unavailable or corrupt | Preserve the historical conclusion and clearly show current verification gaps and integrity state |
| Active task or concurrent writes during export | Use a consistent snapshot or fail explicitly; do not publish a complete bundle assembled from conflicting revisions |
| Report write failure or destination conflict | Preserve existing results and execution references; allow a separate export retry without remeasurement |
| Ordinary Markdown viewer and offline browser | The default entry point is readable, core HTML content needs no network, and external evidence is identified |
Roll out in three steps: define the field contract, adapter rules, and representative examples; implement shared projection and Markdown/HTML export; then integrate automatic final delivery and verify reproduction entry points using real tasks. Acceptance of the delivery feature and performance acceptance of a particular optimization task remain separate.
Decisions requested
Adopt the following defaults, with review comments identifying any proposed changes:
- Make the four entry files the standard delivery for new tasks, with
report.mdas the default reading entry point. - Introduce the separate
campaign-result.jsoncontract while preserving existing summaries and Apps operation output. - Export automatically after final evaluation and allow partial delivery for failed or incomplete tasks.
- Permit external evidence while requiring provenance, integrity, and availability information.
- Support Markdown and offline HTML initially; add other document formats separately.
Existing implementation references
These references identify the existing behavior and distinguish it from the proposed additions. They are pinned to source revision 016c7c7df023944f0f46d8e777d6d31f1df1764c:
- Shared optimization workflow: one control state, retained candidates, independent final evaluation, and rebuildable summaries.
- Optimization benchmarks: result summaries, round evidence, final reevaluation, and usage provenance.
- Dashboard: offline HTML and read-only presentation boundaries.
- Apps output publication: operation output, fixed references, publication failures, and durability limits.
- Optimization controller: current summary fields and state management.
- Apps results and Apps archive: existing operation result contracts and publication mechanisms.
- Evidence tree: the
plan/,results/,patch/, andvalidator/archive structure.
The new files, fields, and export command described here remain proposed until implemented.
Ownership and RFC completion
- Proposing account: amy-why-3459. Drafted and published by Codex at the user's explicit request.
- Authorization: The user requested translation and publication of this RFC. Implementation scope and owners are to be recorded after the proposal decision.
- Issue owner and implementers: To be designated by repository maintainers.
- Decision and acceptance owner: To be designated by repository maintainers.
- RFC completion: The designated decision owner records adoption, requested changes, or deferral of the delivery files, result semantics, and rollout. Accepted implementation work has linked tasks and owners, with feature acceptance assessed against the criteria above.
- Closure authority: The designated decision or acceptance owner, or an explicitly authorized delegate.
Activity
Dong1017 commentedon Nov 16, 2026
Posted by an agent on behalf of @Dong1017 (incremental review only; no prior comments exist on this RFC as of this review).
Supportive of the delivery contract.
Before adoption
artifacts.jsonrelates to the archive authority adopted in [RFC] Tiered Experiment Artifact Archiving and PR Size Control #464. This RFC gives the bundle its ownartifacts.jsonwith anexport_idand entries for reports, source records, evidence, and external artifacts (§§4, 12). [RFC] Tiered Experiment Artifact Archiving and PR Size Control #464's decision makes each campaign'sresults/artifacts.jsonauthoritative for external archive metadata and requires tooling to normalize the existing Archive MiniCPM Seed-TTS details and retain verifiable core evidence #435/docs: record MiniMax-H3-Image Ascend iteration evidence #394 manifest formats through a shared versioned reader (decision). Please state whether the delivery index is that same manifest, a generated projection of it, or a separate bundle index that references it. If both files remain, define which fields are canonical and how IDs, locations, digests, sizes, availability, and integrity are reconciled, so export does not create a second source of archive metadata. The four files can remain the user-facing delivery surface while [RFC] Tiered Experiment Artifact Archiving and PR Size Control #464's evidence bundle remains the archive layout; the relationship between their indexes is the missing contract.Optional
No additional optional points.
Checking scope: read #468 in full, confirmed it currently has no comments, read #464 and its latest decision/comment thread, and checked
origin/main's optimization summary and dashboard/report paths. Relevant references: #468 §§4, 12 and 14;docs/optimization-benchmarks.md(summary and dashboard behavior);docs/dashboard.md(offline snapshot behavior);src/inferrsi/harness/optimization.py(read_recordsrebuilds summary from plan/state). Did not run code or tests because this was a design review.amy-why-3459 commentedon Nov 16, 2026
@Dong1017 @zengchuang-hw
zengchuang-hw commentedon Nov 16, 2026
Independent RFC review in response to the author’s request above. I reviewed the proposal and its pinned implementation baseline
016c7c7df023944f0f46d8e777d6d31f1df1764c, without using other reviewers’ conclusions.Conclusion: I support the delivery-only direction and the separation of workflow completion, candidate selection, and acceptance. This is a design recommendation, not adoption of repository policy or implementation acceptance. Before treating v1 as implementation-ready, I suggest making the following adapter and safety obligations explicit.
Define a degraded read path for damaged historical evidence. The RFC correctly preserves a historical PASS while exposing current verification gaps. However,
optimization._loadandread_recordscurrently reject missing or changed retained evidence before returning records. Simply reusing that reader will fail export rather than produce the promised partial report. Specify a read-only adapter result that separates independently readable historical assessment facts from present integrity failures. Do not weaken the normal execution reader, rerun assessment, or promote a malformed/unreadable assessment to an established historical PASS. Add fixtures for (a) intact assessment but missing measurement, (b) modified measurement, and (c) missing/corrupt assessment itself; only the facts actually recoverable in each case should survive.Specify how a consistent snapshot is obtained, not just that one is required.
read_recordsreads validated records but does not itself acquire the writer’s lock. State updates and the rebuildable summary are separate writes. The adapter should either use a supported snapshot/lock boundary or verify a stable source revision/digest set across the entire read and retry with a bound; exhaustion must fail explicitly. Also freeze the observation cutoff used for live elapsed time:_summaryotherwise calls the current clock for an unsettled task. A concurrency test should interleave settlement with export and assert that every accepted bundle refers to one coherent source snapshot, without relying onsummary.jsonas authority.Keep source semantics in the adapters rather than copying superficially similar fields.
_summarycounts settled round records, reports caller-round usage separately from the final/session observation, and can emit baseline speedup 1.0 from assessment completion. Those fields are not independently sufficient to establish the RFC’s completion, usage coverage, or final-qualification conditions. Include an explicit mapping table and complete golden examples for each source adapter, particularly completed-but-FAIL, baseline retained with missing/failed final, and partially observed usage. Preserve source verdicts and absent identities rather than manufacturing a claim ID or inferring PASS from execution status. The existing Apps primary-claim validation is a useful boundary to retain.Add hostile-input and publication-failure acceptance cases. Offline HTML still consumes untrusted hypotheses, labels, paths, and imported records. Require safe HTML/embedded-data escaping, safe Markdown links, no automatic fetching of arbitrary external references, and traversal/symlink rejection for bundled attachments. Test script-closing strings and unsafe URL schemes. Redacted shared copies need their own content identity and must not silently replace frozen evidence. For publication, test failure after rename but before the final directory sync: the existing publisher can report an error after output is visible, so reconciliation must inspect the same export ID before retrying.
Verification performed: read the RFC, repository AGENTS and privacy rules, workflow/publication/dashboard documentation, and the cited controller, Apps result-validation, and publication code. Checked current Issue state and my own existing comments before posting. Not performed: no exporter implementation exists in this proposal to execute; no unit tests, concurrency/fault-injection tests, browser rendering, or real-task export/reproduction were run. The four points above are proposed implementation acceptance obligations, not claims that those new paths have passed.
david6666666 commentedon Nov 16, 2026
Could this contract distinguish candidate identity from file integrity? The proposed service-reuse mode in #467 excludes file fingerprints, while this RFC says
finalmust bind the same candidate fingerprint. Please explicitly exclude that mode from exports, or allow a retained exact revision plus fixed run and configuration references to establish the binding. If it cannot be verified, preserve the original assessment for audit without marking the delivered code as accepted. A no-fingerprint fixture would make this boundary testable.Dong1017 commentedon Nov 16, 2026
Acting as authorized steward for @Dong1017 (hosted Dong-bot on
usage_limit).On david6666666's fingerprint / service-reuse question: agree this must be explicit before adoption.
#467accepted service-reuse mode may exclude file fingerprints for some retained candidates.finalbinding must not silently invent a fingerprint export for that mode.This is a design constraint for the next RFC revision — still not adoption. Prior open item remains: define delivery
artifacts.jsonvs#464archive authority before anystatus:accepted.No implementation until those two are written into the RFC body.