Strengthen validation retries and checkpoint safety

This commit is contained in:
2026-08-27 11:50:37 +00:00
parent b2e83bd6e7
commit e6b2ae88d2
73 changed files with 934 additions and 1560 deletions

View File

@@ -134,6 +134,13 @@ relatedness validators report advisory evidence concerns. The configured order
is documented in
[Configuration](../config.md#production-validator-keys-and-default-chains).
Every D&D rejection describes the correction in transcript-grounded domain
terms, using contextual names, artifact fields, and source segment ranges when
useful. The guidance must not ask the model to reproduce durable entity IDs,
hashes, validator module keys, or reason codes. Those identifiers remain in
ordinary validation provenance; only the actionable semantic guidance is
eligible for the correction prompt.
Enemy-event extraction additionally rejects a second `engaged` observation for
the same comparison identity within one scene-scoped result. Normalization may
combine results from distinct scenes, so it intentionally does not apply that

View File

@@ -45,6 +45,15 @@ they need, register each leaf implementation, and add any family-owned assets
or default validator chains. They return contextual errors so production
composition fails at startup rather than at the first run.
A validator that returns a completed rejection must supply two separate
bounded values: a stable `ReasonCode` for provenance and actionable
`CorrectionGuidance` for the producer. Guidance identifies the semantic defect
and the constraints on one complete corrected replacement. It must not contain
validator keys, diagnostic paths, opaque application IDs, or other internal
identifiers. An operator-facing `Message` may explain the same event, but the
framework never copies it into a model request. Missing or invalid guidance is
a validator contract failure.
An artifact family can register an optional typed evidence projector alongside
its codec. The projector returns defensive copies of the artifact's direct
generic source references and must use the codec's exact Go type. It does not
@@ -131,6 +140,8 @@ its domain prompt.
4. Register the module through its typed registry helper and add it to the
owning family registrar. Add a default validator chain only when that
family owns the behavior; otherwise require an explicit compatible chain.
Every rejection path in a validator must provide actionable correction
guidance while retaining its stable internal reason code.
5. Update the selectable-key and chain reference in
[Configuration](../config.md#production-module-keys), the applicable
integration contract, and focused tests. Keep the configuration document

View File

@@ -143,7 +143,11 @@ immutable candidate; it does not regenerate the producer or alter the
validator request. Rejections stop that validator, while other configured
validators still run. The executor retains ordered results, bounded
deduplicated correction guidance from rejections, and only the final exhausted
failure outcome for each validator.
failure outcome for each validator. The correction builder keeps first
occurrence order, omits internal reason codes, validator names, and operator
messages, and requests one complete replacement. Missing guidance or an
oversized aggregate is a framework contract error; guidance is never inferred
or truncated.
The runner applies the binding's retry policy around a stage operation and its
complete validation chain. It preserves warnings only from the final accepted
@@ -163,12 +167,14 @@ settles the semantic policy immediately. Structural-output errors alone use the
structural policy, and validation failure without rejection settles the
validator-failure policy without regenerating the producer.
Chunk planning uses this state machine for generated plans. A rejected
automatic cache hit is not model material and therefore falls through to a
fresh initial generation; it neither receives a correction nor overwrites the
stored record. Only a newly generated, completely validated plan is published
to the chunk-plan store. Rejected plans never advance, and validation-incomplete
plans remain unpublishable.
Chunk planning uses this state machine for generated plans. A rejected or
validation-incomplete automatic cache hit is not model material and therefore
falls through to a fresh initial generation at producer attempt one; it neither
receives a correction, consumes retry budget, promotes cached-candidate
warnings, nor overwrites the stored record. An incomplete cache validation
under `fail_run` terminates instead. Only a newly generated, completely
validated plan is published to the chunk-plan store. Rejected plans never
advance, and validation-incomplete plans remain unpublishable.
After terminal lane work, the runner assembles manifest provenance, normalized
artifacts, rejections, warnings, and an optional accepted chunk map. When an
@@ -204,7 +210,12 @@ The runner writes successful checkpoint artifacts only after complete accepted
validation. Chunk plans follow the same rule for publication. A rejection,
invalid structured response, or incomplete validation is never reusable state;
the current run may still hand off an otherwise valid `warn_continue` result
according to its terminal policy. Attempt debug records retain safe kind,
according to its terminal policy. The runner carries private reuse eligibility
through extract, merge, normalize, and generated-reference handoff. Any stage
derived from incomplete validation skips both checkpoint lookup and all
checkpoint publication even when that stage's own validation completes.
External references and fully validated generated references remain eligible.
Attempt debug records retain safe kind,
validator, repair-usage, policy, and terminal-decision provenance. Full
assistant and correction content remains confined to the requested detailed
LLM trace.