AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row

[seathatflowsinourveins]

Scope

• What this PR changes: it records the second-family lanes for AN-13's single-family findings M4 and M5.
  • M4: both lanes reject the report's rewrite of the vendored verification-before-completion skill, so it gets
    no edit. The lanes did not test the skills trial's immediate-removal rule
    (2026-09-25-skills-trial-and-usage.md:284-286); the record says so, leaves that question open and makes a
    two-family finding of conflict an overturn condition.
  • M5 (blueprints/convergence-practice/application-delivery/AGENTS.md:15-18): both lanes amend it, but with
    different texts. One blind adjudication, in both orders by both families, split, so the lines are unchanged and
    both texts are recorded.
  • Anti-pattern row: this round built its judge audit from another round's audit instead of the lane-A template
    that the log names, and dispatched before that template's checks ran. Both Claude judgments were void on false
    positives the template already covers; the outcome does not depend on them.
• Base commit: 3058b237.
• Lane: lane:foundation. The paths are docs/ and evidence/artifacts/, plus the manifest re-registration,
  which does not make a PR shared (docs/lanes.md).
• Owned paths touched:
  • docs/decisions/2026-09-28-an13-m4-m5.md (new);
  • evidence/artifacts/prompt-audit-an13-m45-20260928/ (new, 62 files);
  • one row in docs/harness-defaults.md, placed where no open PR edits the log;
  • manifests/evidence.json, re-registered in the last commit.

SOTA sources

• M4, the pinned upstream skill (kept): obra/superpowers@8ca22dba skills/verification-before-completion/SKILL.md.
  • Its sha256 is 2befe7fc…, the pin's skill_md_sha256 in adoption/skills/manifest.json.
  • Upstream main served the same blob on 2026-09-28.
• M4, upstream's latest change to that file: obra/superpowers commit 3be5aad3dd24
  (2026-07-24). It dropped the persuasion sections and kept the Iron Law, the gate function and the rationalization
  table: "This is the eval-gated class … evals on this branch decide".
• M4, the concern's support: Anthropic, Prompting Claude Opus 5, "Task scope and over-verification",
  retrieved 2026-09-28.
• M5, both lanes' prompting sources:
  • Anthropic, Claude prompting best practices, "Be clear and direct";
  • OpenAI, latest-model guide (GPT-6 Astra and
    AGENTS.md).
• Method:
  • the convergence rule, docs/decisions/2026-09-27-prompt-audit-resolution.md:54-60 ("Rule"). The rule that a
    void judgment does not count is from the lane-A template's outcome-actions.md:9-12 and
    tools/sota-convergence/README.md:1503-1505;
  • the packaged GPT-6 runner, tools/sota-convergence/landscape-sweep/codex_call.sh (codex-cli 0.157.1,
    gpt-6-astra, effort max);
  • the blind-adjudicator agent, .claude/agents/blind-adjudicator.md;
  • the lane-A judge-audit template the log names, evidence/artifacts/prompt-audit-20260927/lane-a/final/adjudication/
    (its self-test and root scan, run afterwards against this round's audit);
  • the anti-pattern log rule in docs/harness-defaults.md, "Record a correction the same turn".

Evidence-class table

Claim	Evidence class	Command / receipt
Both lanes returned reject for M4 (0.91, 0.87) and amend for M5 (0.97, 0.74), with the two texts quoted in the record	native_proven (runs on this host, records retained)	round1/returns/, round1/runner/
M5's adjudication: two valid GPT-6 judgments for different texts; both Claude judgments void; outcome split	native_proven (runs), local_integration (audit and tally)	adjudication/{gpt6,claude}.{AB,BA}.json, m45-mapping.json, audit-m45.json, tally-m45.json
Both Claude voids are false positives: "Codex lane" in the adjudicator's own definition, and evidence-reviewer in a listing of the export's agent files	local_integration	adjudication/audit-m45-windows.json (each hit's window, host details replaced)
The lane-A template's two checks, applied to this round's audit and patterns without the template's own scripts, both fail (the new row's check)	local_integration (reproduction after the fact, repeated byte for byte)	python3 -B scripts/reproduce_template_checks_m45.py <clone> 9f8db582 <empty dir> <out> exits 0; adjudication/template-checks-m45.json
The audit reproduces, and voids planted accesses. It is not the pre-dispatch control: it has no clean-run criterion, so it cannot catch false positives	local_integration (after the judges)	adjudication/audit-control-m45.json
The Claude lane and both Claude judges called only Read, Glob, Grep and one advisor each	local_integration (transcript counts)	tool-counts.json
The packet and the audit's scripts are unchanged since freezing	local_integration	sha256sum -c packet.sha256 in round1/; sha256sum -c frozen-sha256.txt in scripts/
M4's reasons (the pin, upstream's choice, the pinned preload, the scheduled comparison)	source_review	decision record, "M4"
No model behavior changed or was measured	— (nothing claimed)	decision record, "Limitations and residuals"

Local commands run

On a35194e8 (the repair round, a wording fix and the M4 limitation, re-registered on 3058b237). The full suite
ran on 4341285d. After it, c94c87aa and 968e6f9a change only prose, the reproduction script's docstring and
its output's note; e3791344 and a35194e8 re-register.

$ python3 -m unittest tests.test_adoption_docs_consistency        # after the row's rewrite
Ran 37 tests … OK (skipped=1)
$ python3 scripts/validate.py                                      # after re-registration
{"components": 69, "hashed_files": 7726, "profiles": 4, "receipts": 163, "status": "passed"}
$ bash ci_local_fast.sh <worktree> <logs>                          # validate.yml's local steps, in CI order
workflows-sha256, workflows-contracts, zizmor, validate, host-receipts, release-due, validate-catalogs,
validate-foundation, landscape, audit-reports, validate-convergence, build-verdicts, component-matrix,
grand-list, verdict-flip, gap-crosswalk, gap-wave-ledger, build-ecosystem, skill-contract: all exit 0
$ python3 -m unittest                                              # full suite, on 4341285d
Ran 6889 tests in 750.071s; FAILED (failures=1, errors=1, skipped=765)
  ERROR: tests.test_order_throughput.CapacityRunTests.test_cli_refuses_live_base_url_from_env_file_without_network
  FAIL: tests.test_gpt6_family_tiering_20260926.RunnerTests.test_sigterm_records_the_running_call_as_interrupted_without_counting_it
# the same two as on the first head; both fail identically when run alone on a clean detached checkout of
# 3058b237 (unmodified main): this host has a /tmp/.git directory, and both tests refuse temp paths inside a repository
$ python3 -B scripts/reproduce_template_checks_m45.py <worktree> 9f8db582 <empty dir> <out>
{"payloads_void": [true, true], "gpt6_clean": true, "path_hits": 8, "text_hits": 242, "reproduced": true}
exit 0; this run wrote adjudication/template-checks-m45.json, and a repeat of the earlier version was byte-identical
$ (cd evidence/artifacts/prompt-audit-an13-m45-20260928/round1 && sha256sum -c packet.sha256)
packet.md: OK
$ (cd evidence/artifacts/prompt-audit-an13-m45-20260928/scripts && sha256sum -c frozen-sha256.txt)
void_patterns_m45.py: OK / audit_m45.py: OK / make_m5_adjudication.py: OK
$ pre-push scans: added lines and messages for UUIDs, home paths, the host user name, emails, session scratch,
  secret-like strings and agent ids: 0 each; validate.py's PRIVATE_CONTENT patterns: 0 hits in the 64 files;
  client-configuration text in the package: one hit, `.mcp.json` named as a scan category in the review brief,
  with no configuration quoted

Decision record

docs/decisions/2026-09-28-an13-m4-m5.md. It gives:

• the decision per item, both M5 texts verbatim, the method and checks;
• the audit's false positives, the skipped template and its checks run afterwards;
• usage, with each provider's counters separate;
• alternatives, overturn conditions and limitations.

Review

One review per family ran on the local head 31ab17f7, before the first push. One repair round followed: 297709c6,
re-registered in 4341285d. A wording fix found before merge followed: c94c87aa, re-registered in e3791344. Then a gap raised after the
reviews: 968e6f9a, re-registered in a35194e8. Each finding and what happened to it is in
review/README.md, with both prompts, returns
and usage.

• GPT-6: codex exec --json -s read-only -m gpt-6-astra -c model_reasoning_effort="max" -c web_search="live",
  14:48:54Z to 15:03:54Z. Changes-needed: four should-fix findings and two nits, all on the record's wording. They
  covered a lane attribution, the rule's citation, the freeze-time evidence, what the judges' inputs held, a receipt
  citation and a checksum invocation. It confirmed the outcomes, the recomputed tally, the other repository
  citations, the upstream pin and commit, the usage figures, the sent hashes and the registration.

• Claude: one evidence-reviewer (Opus 5.5, effort max), given the command results. Changes-needed: six
  should-fix findings and three nits. The two main ones:

  • the round adapted X9 round 3's audit instead of the lane-A template that the log names, which already covers both
    false positives;
  • the row's "Where enforced" cited a control that cannot fail on this mistake.

  It confirmed the outcomes, that the outcome does not depend on the voids, the figures, the inputs' symmetry,
  privacy and the registration.

• Repairs: every finding was fixed. The row now names the template, and its check is the template reproduction
  above. A check before merge found that the row and the record said the template's checks were run; the
  reproduction applies them without running the template's own scripts, and the wording now says so. The hit windows and tool counts are published. My own check after the repair found one more defect: the
  record cited the void rule to lines that do not state it. It now cites that rule's own sources.

• The M4 gap: the packet did not include the skills trial's rule that removes a trial skill whose instructions
  conflict with CLAUDE.md or AGENTS.md, and neither return mentions it. The record now says M4 covers the report's
  rewrite only, and adds a two-family finding of conflict as an overturn condition.

• Residuals: no second review round ran, so the repairs had no independent review. The adjudication packet is
  kept as sent, with its wrong receipt.json:88 premise; the record's limitations explain it.

Host evidence

Not applicable: nothing under evidence/hosts/ changes.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a
  version comment (no floating tags). No Actions change.
• New/changed workflows declare top-level permissions: contents: read
  (or a narrower, explicitly justified addition). No workflow change.
• No secrets are printed, logged or committed; no new required secret was
  added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved (not deleted,
  moved or overwritten).

🤖 Generated with Claude Code