Transaction-Science/open-standards
Watch
1
0
Fork
You've already forked open-standards
 0
Code Issues Pull requests Projects Releases 15 Packages Wiki Activity Actions

jouleclaw: an answer carries its verifier's backing, so a cached answer is not refused as ungrounded #276

Merged
dcharlot merged 1 commit from fix/answer-carries-entailment-backing into main 2026-10-16 09:41:29 -04:00
Conversation 0 Commits 1 Files changed 3 +173 -18
dcharlot commented 2026-10-16 04:48:15 -04:00
Owner
Copy link

What was wrong

An answer's Provenance stored the entailment results the verifier consulted, but not the backing the verifier declared for them (VerificationReport::grounding_backing).

jouleclaw-edge-cli rebuilds a VerificationReport from a cached answer. Without a stored backing, the rebuild could only declare the floor, Backing::Unstated. As a result, require_substantive_grounding refused every cached answer as ungrounded, including answers a real NLI model had grounded. The rebuild's own comment named the fix: persist the backing beside the entailments.

The fix

  • Provenance.entailment_backing: Option<Backing>. It is omitted from the wire when absent, so answers and caches written before it keep their bytes and deserialize as before. Absent reads as unstated.
  • compose_verified_answer copies the report's backing into the answer.
  • The cache rebuild restores that backing. It never derives one from the cached model_ids, because that would let a stand-in pass as a real model. require_substantive_grounding still checks the backing against the cached results:
    • a fixture stays a fixture;
    • a backing that names a different model from the one whose results are attached is refused.

AnswerOrRefusal gains allow(clippy::large_enum_variant), with the reason in a comment. The new field pushed the Answer variant past clippy's size threshold. There is one value per query, and boxing it would change a public enum for no measurable gain.

Checked

Tests run on the build host:

  • jouleclaw-schema, jouleclaw-compose and jouleclaw-edge-cli pass, including the new tests:
    • the_answer_carries_the_verifiers_backing
    • a_cached_answer_keeps_the_backing_its_verifier_declared
    • a_cached_backing_is_never_promoted_or_trusted_past_its_results
    • the answer round trip, which checks the field is absent from the wire when None and round-trips when Some
  • Clippy reports nothing new on lines this change touches.

Each of these mutations fails a test:

  • the composer dropping the backing;
  • the rebuild going back to the floor;
  • the rebuild deriving SingleModel from the cached model_id, which is the stand-in-as-model path.

🤖 Generated with Claude Code

## What was wrong An answer's `Provenance` stored the entailment results the verifier consulted, but not the backing the verifier declared for them (`VerificationReport::grounding_backing`). jouleclaw-edge-cli rebuilds a `VerificationReport` from a cached answer. Without a stored backing, the rebuild could only declare the floor, `Backing::Unstated`. As a result, `require_substantive_grounding` refused every cached answer as ungrounded, including answers a real NLI model had grounded. The rebuild's own comment named the fix: persist the backing beside the entailments. ## The fix - **`Provenance.entailment_backing: Option<Backing>`.** It is omitted from the wire when absent, so answers and caches written before it keep their bytes and deserialize as before. Absent reads as unstated. - **`compose_verified_answer`** copies the report's backing into the answer. - **The cache rebuild** restores that backing. It never derives one from the cached `model_id`s, because that would let a stand-in pass as a real model. `require_substantive_grounding` still checks the backing against the cached results: - a fixture stays a fixture; - a backing that names a different model from the one whose results are attached is refused. `AnswerOrRefusal` gains `allow(clippy::large_enum_variant)`, with the reason in a comment. The new field pushed the `Answer` variant past clippy's size threshold. There is one value per query, and boxing it would change a public enum for no measurable gain. ## Checked Tests run on the build host: - `jouleclaw-schema`, `jouleclaw-compose` and `jouleclaw-edge-cli` pass, including the new tests: - `the_answer_carries_the_verifiers_backing` - `a_cached_answer_keeps_the_backing_its_verifier_declared` - `a_cached_backing_is_never_promoted_or_trusted_past_its_results` - the answer round trip, which checks the field is absent from the wire when `None` and round-trips when `Some` - Clippy reports nothing new on lines this change touches. Each of these mutations fails a test: - the composer dropping the backing; - the rebuild going back to the floor; - the rebuild deriving `SingleModel` from the cached `model_id`, which is the stand-in-as-model path. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
jouleclaw: an answer carries its verifier's backing, so a cached answer is not refused as ungrounded
All checks were successful
jouleclaw public surface must stay documented / Public-surface audit (jouleclaw-rs) (pull_request) Successful in 1m6s
Details
repository checks / lockfiles, tests in CI, workflow hygiene, spec and section references, public text (pull_request) Successful in 52s
Details
standards workspaces build + test / which workspaces this change touches (pull_request) Successful in 4s
Details
standards workspaces build + test / mesh-rs (mesh/mesh-rs, -p mesh-engine-gguf -p mesh-openai --features mesh-engine-gguf/gguf,mesh-openai/server, mesh-rs, linux-arm64) (pull_request) Successful in 1m7s
Details
standards workspaces build + test / macOS energy backends (pull_request) Successful in 6m2s
Details
standards workspaces build + test / jouleclaw-rs (-nodebug, jouleclaw/jouleclaw-rs, -p jouleclaw-energy --features rapl,nvml, jouleclaw-rs, linux-arm64, -C target-feature=+fp16) (pull_request) Successful in 3m35s
Details
standards workspaces build + test / jouledesk-rs (jouledesk/jouledesk-rs, jouledesk-rs, macos) (pull_request) Successful in 1m9s
Details
07756a6730 
An answer's `Provenance` stored the entailment results the verifier consulted,
but not the backing the verifier declared for them
(`VerificationReport::grounding_backing`). So jouleclaw-edge-cli's cache
rebuild could only declare the floor, `Backing::Unstated`, and
`require_substantive_grounding` refused every cached answer, including those
a single real NLI model had grounded. The rebuild's own comment named the
fix: persist the backing beside the entailments.

- `Provenance.entailment_backing: Option<Backing>`. It is omitted from the
  wire when absent, so answers and caches written before it keep their bytes
  and deserialize as before; absent reads as unstated.
- `compose_verified_answer` carries the report's backing into the answer.
- The cache rebuild restores that backing and never derives one from the
  cached `model_id`s. `require_substantive_grounding` still cross-checks it
  against the cached results: a fixture stays a fixture, and a backing that
  names a model other than the one whose results are attached is refused.

Tests:
- `the_answer_carries_the_verifiers_backing`: compose carries any backing,
  the floor included, unchanged.
- `a_cached_answer_keeps_the_backing_its_verifier_declared`: a single model's
  labels are grounded after the round trip.
- `a_cached_backing_is_never_promoted_or_trusted_past_its_results`: fixture
  stays refused, a mismatched model is refused, and an old entry is unstated.
- The answer round trip: the field is absent from the wire when None and
  round-trips when Some.

`AnswerOrRefusal` gains `allow(clippy::large_enum_variant)` with its reason.
There is one per query, and boxing would change a public enum for no
measurable gain.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
Transaction-Science/open-standards!276
No description provided.