[kikashy]

Closes #170. Closes #169.

What it adds

• Declares six closed per-target conversion kinds, their explicit rules and recorded failures.
• Declares precedence, agreement and first-present reconciliation without implicit last-writer-wins.
• Surfaces missing, wrong-type, out-of-domain and conflict findings as unknown facts and records every transformation in preparation lineage.
• Adds the typed research results contract, explicit observation states, versioned acquisition-source profiles, resolved execution settings and retained bytes for offline replay.
• Updates RFC 0018, the dialect companion, both verifier implementations, the changelog and shared hand-written conformance cases.

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 73s
Ran 232 tests in 72.820s
OK
PASS policy-prototype 1s
Ran 38 tests in 1.110s
OK
PASS policy-contract 2s
Ran 13 tests in 2.427s
OK
PASS policy-official-draft 2s
Ran 16 tests in 2.226s
OK
== 7 passed, 0 failed, 0 skipped, 0 not run ==

Not in this PR

Runner implementation remains tracked by Runner #86, and gateway acquisition capabilities remain tracked by Gateway #275.

A Codex session wrote this change.

[kikashy]

Review round 1 (OpenAI gpt-5.6-sol, effort medium, 2026-10-23), reviewed SHA 336b354

Clean-room clone at the head, probes under .review/; CI green before the round (the head since carried past main, 25606ba, with the same content). The report verbatim:

HIGH — tools/dialect_verify.py:2477: readers disagree on date width. For MM/DD/YYYY with "1/2/2026", Python accepts 2026-01-02, while TypeScript’s exact-width regex at tools/dialect-verify-ts/src/prepare.ts:120 records invalid-encoding. Fix Python to enforce the declared lexical form before calendar parsing and add an adversarial vector.

HIGH — tools/dialect-verify-ts/src/prepare.ts:84: decimal-from-number diverges for exponent literals. For 0.01e3, Python produces "10"; TypeScript constructs "0010", then records wrong-type and omits the fact. Fix and specify one exact exponent-expansion/zero-normalization algorithm; add exponent-shift and negative-zero vectors.

HIGH — spec/dialect.md:1066: decimal-from-text does not define valid thousands-separator placement. Both readers currently erase every occurrence, admitting "1,,2" as "12", while another text-conforming reader could reject it. Define the grouping grammar—or explicitly define unconditional replacement—and vector malformed grouping.

HIGH — tools/dialect_verify.py:1914: typed results admits legacy found and not-found states; TypeScript does likewise at tools/dialect-verify-ts/src/forms.ts:1415. Both probes were accepted although §4.3 closes typed results to present, missing, conflicting, inaccessible, and stale. Branch status validation by contract spelling and add rejection vectors.

HIGH — spec/dialect.md:1094: the text says every value outside its declared type records wrong-type, but line 1099 and both readers treat null as valid for every quality type and write it into facts (tools/dialect_verify.py:2487). A text-only implementation can legitimately omit it as unknown instead. State one outcome explicitly; to preserve the typed target contract, record wrong-type, omit the target, and vector it.

MEDIUM — tests/test_dialect.py:2731: the research “replay” vector only calls form checking and copy_claim; it never enters the public verifier or constructs a preparation. retainedBytes is only base64-validated at tools/dialect_verify.py:1982, so changing it to unrelated valid bytes leaves acceptance and results unchanged. Add a complete runnable verifier vector with expected preparation/lineage and exercise it through both verifier entry points.

LOW — rfcs/0018-dialect.md:953: the amendment record calls execution evidence optional, while §4.3 and both readers require execution for results. Say “required for results, optional for legacy fields.”

Verified

• Reviewed all 11 changed files at 336b354a1bc341251a57b486859b7a1e6e784cdc.
• git diff --check origin/main...HEAD passed.
• Python changed-file suite: 82 tests passed.
• Hand probes covered five conversions and all reconciliation modes; the supplied precedence, first-present, agree, and conflict answers matched their vectors.
• Digest and lineage integration is present in both readers; last-writer-wins is rejected.
• No normative text requires a model or web call during replay.
• TypeScript was inspected for equivalent checks, but its suite could not be rerun because the available Node is v10.19.0 and lacks node --test; CI was already reported passing.

merge: after fixes — 336b354

Disposition (the coordinator's)

#	Finding	Severity	Disposition
1	The two readers disagree on date width (1/2/2026 under MM/DD/YYYY)	HIGH	Fix in this pull request by the same builder: the declared lexical form enforced before calendar parsing in both readers; a vector
2	decimal-from-number diverges on exponent literals (0.01e3)	HIGH	Fix in-PR: one exact exponent-expansion and zero-normalization algorithm in the text, both readers; exponent-shift and negative-zero vectors
3	decimal-from-text leaves thousands-separator placement undefined (1,,2 admitted)	HIGH	Fix in-PR: the grouping grammar defined in the text, both readers; a malformed-grouping vector
4	Typed results admit the legacy found/not-found states	HIGH	Fix in-PR: status validated by the contract spelling in both readers; rejection vectors
5	null outside a declared type: the text says wrong-type, the readers write it as valid	HIGH	Fix in-PR: one outcome stated, wrong-type recorded and the target omitted, both readers; a vector
6	The research replay vector never enters the verifier; retainedBytes only base64-checked	MEDIUM	Fixed in the same push: a complete runnable verifier vector with expected preparation, lineage and execution record; changing the retained bytes changes the result
7	The amendment record calls execution evidence optional for results	LOW	Fixed in the same push

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

[kikashy]

Re-read after the fix (OpenAI gpt-5.6-sol, effort medium, 2026-10-23), reviewed SHA e51da60

The same reviewer, clean-room clone at the fixed head with Node 22 for the TypeScript verifier, probes under .review/; CI green. The report verbatim:

Findings: none.

Verified

• All five HIGH fixes align across normative text, Python, TypeScript, and shared vectors.
• Scratch reversions failed date-width, exponent negative-zero, and malformed-grouping as expected.
• Typed results reject legacy states; null produces wrong-type.
• Both public verifiers accept retained bytes and refuse the changed-byte twin with preparation-differs.
• The amendment now requires execution evidence for results.
• Checks: Python 83/83 passed; Node 22 fix-focused TypeScript tests passed.

merge: yes
reviewed SHA: e51da60fde9e887173a6e6d59ff69160d1c544df

Disposition: all five HIGHs, the MEDIUM and the LOW are fixed; no further round. Merged on the standing rule. Same vendor as the builder, not a cross-vendor claim.

[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