Release Acceptance

An index change is releasable when a reviewer can connect execution intent to backend capability, ranked output, provenance, and replay behavior. Returning plausible neighbors is not acceptance evidence by itself.

flowchart LR
    request[Execution request]
    plan[Immutable plan]
    backend[Declared backend]
    result[Ranked result]
    ledger[Artifact and provenance]
    replay[Replay verdict]

    request --> plan --> backend --> result --> ledger --> replay

Acceptance record

Changed surface	Required evidence	Release-blocking result
scoring, metric, or tie order	focused scoring fixtures and cross-backend exact conformance	equal inputs inherit incidental backend order
execution request or plan	ABI, immutability, normalization, and fingerprint comparison	materially different intent shares an execution identity
exact backend	capability declaration plus deterministic CRUD/query conformance	an exact claim produces unstable ranked output
ANN backend	exact baseline diff, randomness record, parameters, seed behavior, and witness	approximation is presented without its loss or randomness boundary
budget enforcement	refusal and partial-result scenarios for each budget dimension	a breached budget is reported as complete success
artifact or run lifecycle	incomplete, failed, complete, corrupt, migration, and portability fixtures	an unfinalized or corrupt run loads as complete
replay	stored baseline, provenance join, golden comparison, and changed-input refusal	drift is relabeled as a match or omitted
public boundary	DTO, authorization, idempotency, schema, and error-contract evidence	HTTP or CLI output changes domain meaning

Keep execution verdicts separate

Vector execution crosses several independently reviewable decisions. A release record must preserve the strongest state actually reached:

State	Required evidence	Stronger claim still unavailable
admissible	normalized request, intent, mode, contract, budget, and capability match	the backend executed the request
executed	backend identity, parameters, input fingerprint, ranked output, cost, warnings, and refusal status	the result was durably finalized
finalized	complete artifact lifecycle, provenance ledger, fingerprints, and native-state reference	replay produces an acceptable comparison
replay compared	original and candidate envelopes, policy, structured diff, verdict, and reason	equivalence outside the declared policy
portable	isolated reader or import test plus every referenced external snapshot or durable identifier	another backend will rank identically

An ExecutionArtifact in a complete state proves only what its envelope and provenance retain. If the native ANN index, remote collection, embedding model, or dataset snapshot is external, portability and replay remain conditional on that external identity being available.

Release comparison set

Select fixtures that expose the algorithmic and lifecycle edges affected by the change:

1. one exact request with ties, stable ordering, and a known result;
2. one bounded or approximate request compared with its exact baseline;
3. one unsupported-capability or budget refusal;
4. one incomplete or corrupt artifact that must not load as complete;
5. one replay with no disallowed drift and one with a single blocking drift;
6. one alternate eligible backend for the shared conformance boundary; and
7. one public-interface request whose error retains the domain failure class.

The exact, ANN, conformance, provenance, execution-diff, misuse, and API/CLI suites each defend a different part of this set. Report which fixture supports which claim; a large passing count without request, backend, and artifact identity is not an interpretable release record.

Backend-specific custody

Retain the dataset and vector identities, request, immutable plan, backend name and version, capability descriptor, parameters, randomness profile, budget, ranked output, provenance ledger, and lifecycle state. External ANN indexes or remote databases require their own snapshot or durable identity; a portable JSON record does not contain those systems.

Release decision

• Exact evidence supports equality only for the recorded contract and numerical environment.
• ANN evidence supports declared approximation and replay interpretation, not exact equality.
• Conformance supports a shared protocol, not identical ranking across every backend.
• Performance evidence is comparable only with dataset, backend, parameters, dependency versions, and hardware held visible.

Use change validation for evidence routing and known limitations for the boundaries that remain after acceptance.