test: remove the AI scorer that grades skill prose for clarity (freeze repair, GST-1371) (#4009)

* test: remove the skill-prose clarity LLM judges (GST-1371)

Removes the 21 eval cases where an LLM scored a skill's instruction text for
clarity, completeness and actionability, with their registrations, input
builders, response cache, calibration corpora, the cookie-workflow manual-review
approval and scripts/judge-controls.ts. The three remaining skill-llm-eval
judges keep their prompts, thresholds and samples; the file's supervision wall
shrinks to fit them. Outcome judges are untouched.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test: pin the retired judge-control script and cookie manual-review record as deleted (GST-1371)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test: register the manual-review fixture as free (GST-1371)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test: keep the orphaned clarity-judge calibration corpora for a follow-up deletion (GST-1371)

Deleting the four corpus directories alone puts the diff over the /ship review
ceiling; they serve no configuration now and go in a follow-up.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* docs(todos): file the clarity-judge removal follow-ups (GST-1371)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

* test: keep skill-bundle section paths POSIX so references dedupe on Windows (GST-1371)

windows-scoped-shard (run 41918796523) failed 'cross-skill references retain
complete bytes once': path.join gave the section entry a backslash path, so a
reference to the same section was not recognized as already bundled.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

---------

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
This commit is contained in:
Garry Tanandcapy-ai[bot] authored and GitHub committed 2026-11-04 22:19:08 +00:00
1 parent 622e8b1f76
commit 781f928ccf
55 files changed
+209 -3000

No files matched your search

+10 -18
View File
@@ -324,13 +324,10 @@ changes are explicitly deferred, not counted as PR passes. Weekly/manual runs of
runs the broad gate. This deliberately moves some defect detection later.
`scripts/eval-input-cache.ts` accepts only complete, clean, first-attempt passes
with matching before/after inputs. The audited workflow-judge adapter hashes the
actual expanded prompt, source/fixture/rubric/runner closure, installed SDK,
model parameters and runtime. Missing/unknown inputs force execution. Receipts
with matching before/after inputs. Missing/unknown inputs force execution. Receipts
are scoped to the same repository and PR, expire after 24 hours, and contain
public scores and provenance rather than prompts or secrets. Of the 17 cases
using `runWorkflowJudge`, 16 are eligible; the cookie workflow's custom input
does not match the cache adapter and stays fresh, as do the other 11 quality cases.
public scores and provenance rather than prompts or secrets. The quality judges in
`test/skill-llm-eval.test.ts` never reuse a receipt and always run fresh.
CI supplies the scoped cache/runtime configuration; local runs are fresh by
default. Cached scores must
pass current assertions; reused records retain their original source and time
@@ -455,7 +452,7 @@ and the kind fixes its trials before the run:
non-JSON, a malformed field) fails the panel and is never resampled; a
refusal counts as an unscored panel only when every sample refused.
`callJudge`'s 429 backoff happens before any model output and is transport,
not a verdict retry. The workflow-judge cache stores whole panels only.
not a verdict retry.
The `rule` case auq-matrix uses the same panel inside its one execution: each
skill's single native capture has its recommendation substance scored by 3
judge samples, gated on their median against the unchanged minimum of 4, and a
@@ -840,11 +837,8 @@ cleanup. No per-case budget grows. Overlay wrappers
have a 1,830-second minimum shard wall and run without Bun retries; see the
[overlay contract](OVERLAY_BENCHMARK_CONTRACT.md) for their unchanged work budget.
The quality file reserves its whole-file wall (3,170 seconds) for every case run
once, plus cleanup. Each still has 120 seconds of model work. Its 17 workflow
judges own their deadline and abort signal, with five seconds for terminal
recording inside a ten-second Bun grace; the other 11 retain their existing
120-second Bun timeout. Late responses cannot create records or cache passes.
The quality file reserves its whole-file wall (480 seconds) for every case run
once, plus cleanup. Each of its three judges has a 120-second Bun timeout.
When a judge's abort signal cancels its request, `callJudge`
(`test/helpers/llm-judge.ts`) prints one `llm-judge-aborted` JSON line to
stderr: elapsed time, wrapper `attempts` and `retryWaitsMs` (429 waits),
@@ -1444,7 +1438,7 @@ once and reused from there. It prints:
Screening figures are labelled `screen-provisional`: the rule's preamble
detector is a line-commonality heuristic, not T2's hunk classifier, so they
stay heuristic until T2 lands. A `package.json` change that only moves the
version is neither selected nor a canary trigger, and all 24 judges always run.
version is neither selected nor a canary trigger, and all 3 judges always run.
The selectors are this checkout's, so an old merge is replayed against today's
maps. A shallow checkout says which `git fetch --shallow-since` to run. Logic:
`lib/gate-cost.ts`; tests: `test/gate-cost.test.ts`.
@@ -1470,11 +1464,9 @@ integration tests, the Aside contract pins, and the render-wrapper pins.
`bun run test:pr` runs the selected short live behaviors and quality judges.
It reports deferred broad coverage; unknown dependencies restore the full gate,
and an unmapped prompt without registered coverage blocks planning. Full free
acceptance and required PR checks must pass before publishing. CI can reuse the
16 workflow-judge passes for 24 hours when their complete consumed inputs and
runtime match; records preserve original provenance. The cookie workflow's custom
input, the other 11 judge cases, dynamic agent tests, and local runs without
scoped cache configuration stay fresh.
acceptance and required PR checks must pass before publishing. The quality
judges, dynamic agent tests, and local runs without scoped cache configuration
stay fresh.
Scheduled/manual full coverage and `test:release` always run fresh.
See [testing policy](../CONTRIBUTING.md#test-tiers) for commands and measured targets.
Anything that needs Aside
+1 -1
View File
@@ -15,7 +15,7 @@ different things even when they mention the same skill.
| --- | --- | --- |
| Selection, budgets, process ownership, cleanup, native-event parsing and recorded failure controls | Free runner and fixture regressions | A live model choosing or completing the right workflow |
| Generated files, host parity, section manifests and complete input identity | Free generation and structural tests | An agent actually loading the required section |
| Prompt clarity and rubric quality | The registered quality judge for that complete prompt | Tool execution, native acknowledgments or a finished report |
| Targeted prompt properties (QA anti-refusal, cross-skill consistency, voice) | The registered quality judge for that prompt | Tool execution, native acknowledgments or a finished report |
| Native first-question format and substance | Live SDK/PTY question capture | An answered question or a completed workflow; receipts explicitly say `workflowCompleted: false` |
| Stochastic consistency and verbose/carved comparison | Independent captures, with separate stability and A/B oracles | One successful sample reused as three trials, or one prompt version standing in for the other |
| Decisions, findings and report completion | Per-skill native workflow fixtures | The first question alone, screen text without native evidence, or a generic question count |
+2
View File
@@ -1,5 +1,7 @@
# Workflow-judge control bundles (C1), pre-registered 2026-10-04
> Historical record. The workflow clarity judges and `scripts/judge-controls.ts` were removed in GST-1371 (2026-11-04); the evals that check real outcomes replace them.
Purpose: learn whether the workflow judges that produced low samples in the eight
wave censuses (37174266054 … 37198445662) can separate a sound workflow from a
broken one. Score compression at 4 is a hypothesis; these controls test it.