fix(discovery): state a research run that accepts no claim

[kyle-sexton]

Summary

Outcome-gate rows 7 and 12 of /discovery:research are quantified over accepted claims, so an artifact that lists every claim as a Gap (or as an unresolved Conflict) passed them with zero accepted claims and read like an answer. Row 4 said "every claim", which contradicted the Gap route.

Fix

Option (b) from the issue:

• Row 4 is now quantified over accepted claims.
• New verifier-owned row 14: the index frontmatter's accepted: counts the claims neither listed under Gaps nor left unresolved in Conflicts (a Conflicts claim counts only when its entry resolves in the claim's favor; a refuted finding or a criterion-12 failure filed there does not, matching discipline.md). At accepted: 0 the Summary opens with Inconclusive: no claim accepted. A zero with that line passes; a missing or wrong count, or a zero without the line, fails. The verifier owns it because the issue's acceptance names research-verifier.md as the grader and the final accepted set is only known after rows 4, 7 and 12 are graded.
• On the N-topic path the synthesized root index goes to the verifier for every verifier-owned row (4, 7, 12, 14), not criterion 12 alone, in dispatch.md, research/SKILL.md and research-deep/SKILL.md. The verifier may read the sub-slice indexes a synthesized root names, and grades only the root.
• artifact-shape.md defines accepted: (an integer on every run; the inconclusive line opens the Summary). The researcher sets it in its final write. Every list of the verifier rows (dispatch, gotchas, parent contract, research-deep, evals) now names 14.
• Three new eval cases cover an artifact whose every claim is a Gap, one whose every claim is an unresolved Conflicts entry, and a count that includes a Gap claim. research-deep eval 1 expects the four-row synthesis check.
• SKILL.md's Effort sentence still sits inside the first 20,000 bytes that are re-attached after compaction (it ends at byte 19,992).
• discovery 0.28.4.

Verification

• contract.test.sh run against origin/main's discovery plugin: 19 new or updated pins fail; with the review-fix pins on this PR's previous head, 13 fail. On this head: all contract assertions pass.
• Eight of nine discovery *.test.sh suites exit 0; build-view.test.sh fails only its three headless-browser checks, which fail identically on origin/main in this environment. scripts/validate-plugin-contracts.test.sh PASS=173 FAIL=0.
• check-changelog-parity.sh --check, --check-order, --check-bump origin/main and --check-preserved origin/main pass.
• shellcheck clean; typos clean on the changed files; markdownlint-cli2 0 issues on the changed markdown.

Related

Closes #5833
Related: #5805, #5827

  

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-04T02:10:25.254053Z	6f0ecd6	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[claude]

Claude finished @kyle-sexton's task in 51s —— View job

---

Security review complete ✅

Task: Run /review:security-review (per custom instructions) against PR #6147 at HEAD 6f0ecd6.

• Gather context (PR diff, changed files)
• Run /review:security-review for PR fix(discovery): state a research run that accepts no claim #6147 at HEAD 6f0ecd6
• Post findings

Scope

All 12 changed files are plugin documentation/spec and test assets for the discovery plugin's /discovery:research skill: agent prompt bodies (agents/research-verifier.md, agents/researcher.md), skill/context markdown (SKILL.md ×2, artifact-shape.md, dispatch.md, gotchas.md, reference/parent-contract.md), a bash contract-test script, an eval fixtures JSON, the plugin manifest, and the changelog. No application code, no network/file I/O, no shell execution logic, and no secret/credential handling changed.

Findings

None. This PR only tightens the outcome gate:

• Row 4 is narrowed from "every claim" to "every accepted claim," closing a vacuous-pass loophole rather than opening one.
• New row 14 adds a verifier-owned check (accepted: count vs. actual accepted claims, with a required Inconclusive: no claim accepted. line at zero) — a net-new control, not a removed one.
• scripts/contract.test.sh changes are additional assert_present/assert_absent regex pins against static file content — no shell injection surface (no untrusted input reaches these assertions; patterns are hardcoded).
• evals/evals.json additions are fixture prompts/expectations for the eval harness, not executable logic.

I checked this diff under the instruction-surface deletions lens (docs/conventions/instruction-exception-register/README.md is present in this checkout and was consulted): nothing here deletes, narrows, or softens a standing guardrail. Every change either adds a new gate criterion (row 14) or re-scopes an existing one to be more precise about what it grades — there's no case where a previously-enforced check becomes unenforced.

No GitHub Actions files were touched, so zizmor's advisory lane is not implicated.

[claude]

Claude finished @kyle-sexton's task in 4m 12s —— View job

---

Code review complete

• Gather context (PR diff, changed files, CLAUDE.md/REVIEW.md)
• Run skip-gate checks (PR open, not draft, meaningful change → full review)
• Review changed files for correctness/maintainability
• Post findings

Scope: all 12 changed files under plugins/discovery/ (agents, skills, reference, scripts, evals), diffed against origin/main. Security findings are out of scope for this lane per REVIEW.md (claude-security-review.yml exists in this repo).

Findings: none that clear the high-signal bar.

What I checked and verified directly (reading the files, not just the PR's claims):

• The core fix is coherent: row 4 in skills/research/SKILL.md now reads "every accepted claim," matching rows 7/12, resolving the contradiction with the Gap route the issue described.
• New row 14's text, Owner (verifier), and FAIL action are consistent between skills/research/SKILL.md, agents/research-verifier.md's dispatch brief and "Return exactly this" block, agents/researcher.md's withheld-criteria list and verification_request.criterion string, reference/parent-contract.md, skills/research/context/dispatch.md, and skills/research/context/gotchas.md.
• skills/research/context/artifact-shape.md's new accepted: definition ("claims the evidence table presents that the Gaps section does not list") lines up with how Conflicts-flagged claims (still in the evidence table) vs. Gap claims (excluded) are described elsewhere in SKILL.md.
• Ran a literal grep across plugins/discovery/**/*.{md,json} for stale "rows 4, 7 and 12" / "criteria 4, 7 or 12" style references the new contract assertion (contract.test.sh section 22, assert_absent 'no stale three-row verifier count') is meant to catch — no hits outside the intentionally-scoped mentions (gotchas.md's "Criteria 4, 7 and 12 ask the run to judge..." bullet, which is still true of those three specifically and is followed by a separate clause adding criterion 14; CHANGELOG.md, excluded by the test's own design since it's a historical record).
• Walked through both new eval cases (every-claim-a-gap-is-stated-inconclusive, accepted-count-excludes-gap-claims) against the updated gate/verifier text; the expected outcomes match row 14's pass/fail conditions as specified.
• Confirmed the version bump (0.28.3 → 0.28.4) and CHANGELOG.md entry are present and dated, and the new contract assertions (contract.test.sh section 22) target the files/strings actually changed.

Not independently re-run here (no shell execution available in this review environment): contract.test.sh, the other *.test.sh suites, check-changelog-parity.sh, shellcheck, markdownlint. Those results are author-claimed in the PR description and unverified by me.

[github-actions]

claude-security-review has reviewed this pull request through 6f0ecd6; a later push is reviewed from there.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6f0ecd6881

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

claude-review has reviewed this pull request through 6f0ecd6; a later push is reviewed from there.