Evidence Release Acceptance¶
A reasoning change is releasable when its claims, support, trace, verification, bundle, and replay record constrain one another. Fluent output, a passing process, or a stable fingerprint cannot substitute for that chain.
flowchart LR
spec[Problem specification]
plan[Content-addressed plan]
trace[Ordered trace]
support[Evidence and support spans]
verify[Verification report]
bundle[Manifested bundle]
replay[Frozen replay]
spec --> plan --> trace --> support --> verify --> bundle --> replay
Acceptance record¶
| Changed surface | Required evidence | Release-blocking result |
|---|---|---|
| model, identifier, or canonicalization | validation, stable-ID, cross-platform serialization, and changed-content comparison | different content reuses an identity |
| plan topology | deterministic DAG fixtures, cycle and missing-dependency refusal | an invalid graph reaches execution |
| tool or lifecycle event | call/return linkage, event order, failure, and unfinished-step cases | an orphan or incomplete action is finalized |
| evidence registration | governed path, retained bytes, content digest, span bounds, and chunk identity | support resolves outside the bundle or against changed bytes |
| claim construction | support edges, status, confidence, type, and insufficiency cases | a derived unsupported claim is finalized |
| verifier | one positive and targeted negative case for every affected check | corruption passes or fails only through an unrelated parse error |
| manifested run | core-file inventory, digests, runtime identity, checksum, and incomplete bundle refusal | individually plausible files describe different runs |
| replay | original checksum, frozen tool results, pinned retrieval artifacts, fingerprint, and structured diff | replay consults live retrieval or accepts provenance drift |
| behavioral claim | named corpus, cases, constraints, insufficiency policy, and metric artifacts | artifact validity is presented as scientific usefulness |
Claim disposition must remain explicit¶
Kind, support, verification, and status are independent dimensions. A release must preserve their combination rather than reducing it to confidence:
| Claim condition | Acceptable disposition | Release-blocking distortion |
|---|---|---|
| observed claim with exact retained bytes | proposed or validated according to applicable checks | describing observation as independently verified truth |
| assumed claim declared by the problem or policy | proposed with assumption identity visible | presenting the assumption as retrieved evidence |
| derived claim with a complete support path | proposed, validated, or rejected with findings | removing intermediate support or retaining only final prose |
| support span or digest mismatch | rejected or explicit verification failure | keeping validated status because the surrounding sentence is similar |
| applicable check cannot run | insufficient evidence or non-passing verification result | silently omitting the check from the report |
| relevant evidence is absent | insufficient_evidence where policy permits refusal |
inventing support through confidence, provider reputation, or prior output |
A rejected claim can be evidence that the verification boundary worked. An
insufficient_evidence result can be a correct reasoning outcome. Acceptance
concerns whether the record represents those decisions honestly, not whether
every case ends with a positive conclusion.
Bundle acceptance procedure¶
Review a release fixture as a closed evidence object:
- recompute the specification, plan, claim, evidence, and trace identities from canonical content;
- validate plan topology and match every execution event to an authorized node, call, or return;
- resolve every
SupportRefagainst the retained evidence bytes and recompute its snippet digest; - confirm the verification report contains every applicable check and the intended positive or negative result;
- recompute the manifest inventory and invariant checksum over the exact run files; and
- replay from recorded results only, then retain the structured comparison.
The planning, execution, retrieval, reasoning, trace, verification, CLI, API, and replay suites protect different links in this procedure. A release claim must name the links it exercised. Passing serialization tests alone does not establish support integrity, and a replay fingerprint alone does not establish claim validity.
Custody and interpretation¶
Retain the specification, plan, runtime descriptor, trace, evidence bytes, claims, verification reports, run metadata, manifest, and replay output as one review unit. A trace fingerprint establishes identity of the trace file; it does not establish source authority, entailment, completeness, or truth.
An insufficient_evidence result is acceptable when the case and policy allow
refusal. Hiding that refusal behind an empty or confident answer is not.
Use change validation to route focused evidence and known limitations to bound the resulting release claim.