[kikashy]

Closes #228.

What it adds

• contributingRule on workflow output envelopes, outside Core's unchanged disposition
• one closed summary and one reproducible seeded sample per configured stage
• verifier recomputation and refusal for summary or sample disagreement
• drift as a non-verdict comparison of two verified bundles of the same workflow
• hand-written fixture, mutation, refusal, and resume cases in the RFC 0020 recording plan
• workflow text-contract coverage and an unreleased changelog entry

Checks

== check summary (dev.json) ==
handover: boundary 2 paths, shared 3 paths; see AGENTS.md "Before you hand over"
PASS local-paths 0s
PASS agent-docs 0s
sync-agent-docs: CLAUDE.md matches AGENTS.md
PASS diff-check 0s
PASS repository-tests 105s
Ran 211 tests in 104.055s
OK
PASS policy-prototype 1s
Ran 38 tests in 1.290s
OK
PASS policy-contract 3s
Ran 13 tests in 3.453s
OK
PASS policy-official-draft 3s
Ran 16 tests in 3.353s
OK
== 7 passed, 0 failed, 0 skipped, 0 not run ==

Not in this PR

• the Runtime evaluation record (judgment-pack-runtime#305)
• the Runner executor (judgment-pack-runner#170)
• the Runtime and Runner benchmark work (judgment-pack-runtime#306 and judgment-pack-runner#172)
• a recorded workflow fixture, workflow implementation, or conformance claim

A Codex session wrote the change.

[kikashy]

Review round 1 (OpenAI gpt-5.6-sol, effort medium, 2026-10-22), reviewed SHA 0f363df

Clean-room clone at the head, probes under .review/; CI green before the round. The report verbatim:

Findings

• HIGH — spec/workflow.md:591: The drift result defines only its top-level members; the element forms for stages and records are unspecified, so implementations can emit incompatible field names and structures while satisfying “carries those counter deltas.” Define exact closed forms for stage-delta and changed/added/removed record rows, including nested byRule deltas and null sides.

• HIGH — spec/workflow.md:623: The decision requires schemas and a §11 fixture carrying every new member, but HEAD contains no workflow schema or recorded fixture and explicitly says none is recorded; consequently the required disposition digest and summary recomputation cannot be verified. Add closed workflow schemas and a hand-authored fixture containing all new forms, summaries, samples, refusals, and independently fixed digests.

Verified

• Read the complete origin/main...HEAD diff and companion/RFC correspondence.
• Checked contributing-rule placement, summary counters, canonical disposition byte rules, sample seed/size/ordering/algorithm, verifier refusals, recovery state, and drift semantics.
• Ran python tests/test_workflow.py under ulimit -v 4000000 and timeout: 6 passed. The first package-style invocation loaded zero tests; the corrected direct invocation passed.
• Ran .review/probe_bulk_contract.py under the same limits: found zero workflow fixture records and summaries; fixture digest comparison was unavailable. Confirmed the empty-disposition digest as sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945.
• Ran git diff --check: clean.
• Read the changelog and RFC index; both accurately state that only a fixture plan exists and no fixture is recorded.
• No tracked files changed; only .review/ contains the probe.

merge: after fixes — 0f363df

Disposition (the coordinator's)

#	Finding	Severity	Disposition
1	The drift result's element forms for stages and records are unspecified	HIGH	Fix in this pull request by the same builder: closed forms for the stage delta and the changed, added and removed record rows, nested byRule deltas and null sides included
2	No workflow schema and no recorded §11 fixture carry the new members, so the digest and the recomputation cannot be verified	HIGH	Fix in this pull request: a closed workflow schema and a hand-authored two-stage fixture with the summaries, the sample, the refusals and an independently computed digest; the gap predates this change, but the decision requires every new member carried, so it lands here

The same reviewer re-reads the fixed head. Same vendor as the builder, not a cross-vendor claim.

[github-actions]

Carry needed from base commit a90dfc8.

[kikashy]

Re-read after the fix (OpenAI gpt-5.6-sol, effort medium, 2026-10-23), reviewed SHA 6c141a1

The same reviewer, clean-room clone at the head (the fix plus the carries of #236 and #238), probes under .review/; CI green. The report verbatim:

[MEDIUM] The five sample-refusal vectors are not single-fault cases. Each records selected: [] (for example, conformance/workflow/vectors/sample-seed-empty.json:12), although the offered set contains one outcome record. After repairing only the bad seed or oversized size, the specified algorithm selects digest sha256:7171…7171, leaving a separate sample-mismatch. Update selected in all five vectors to the recomputed selection.

Verified

• Drift stage/record forms are closed, typed, bounded, and define nested byRule deltas and null sides.
• The drift fixture validates and independently recomputes exactly from both bundles.
• All schema object forms use additionalProperties: false; the requested decision members are represented.
• All four summaries, including canonical disposition digests and counters, independently recompute exactly.
• The recorded sample independently redraws exactly.
• Digest and sampling text is deterministic enough for independent implementations.
• Focused workflow tests: 16 passed.

merge: after fixes
reviewed SHA: 6c141a139ca5c632edf9aedfa0be4b2d4fede1e2

Disposition (the coordinator's)

#	Finding	Severity	Disposition
1	The five sample-refusal vectors are not single-fault: selected: [] leaves a second mismatch once the named defect is repaired	MEDIUM	Fixed in this pull request by the same builder: selected set to the digest the specified algorithm draws, so each vector fails for its stated reason only; no further round

Both round-1 HIGHs are confirmed fixed (closed drift forms; schema and recorded fixture recompute exactly). The merge follows CI on the fix. Same vendor as the builder, not a cross-vendor claim.

[kikashy]

The re-read's MEDIUM is fixed at b55f8e4 (the five refusal vectors single-fault; the test proves repairing the named defect alone passes); CI green on the carried head 2889dd7; no further round. Merged on the standing rule.

[github-actions]

🎉 Thank you for your contribution to Judgment Pack!

If you’d like another task, choose a path that interests you:

• Go runtime
• Judgment desk
• Gateway and provenance
• Python and evaluator experiments
• Specification and documentation