Prompt audit rounds 2-3: AGENTS.md F1, F4, S1 and X5b; X9 applied to CLAUDE.md, 4 of 4 blind judgments; dashboard manifest bound

[seathatflowsinourveins]

Scope

• What this PR changes: the 2026-09-27 prompt audit's instruction-file items that converged across both model families, and the evidence for each. The co-change the evidence registration needed, the dashboard's read bound, is now on main through Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446.
  • AGENTS.md:
    • F1 (line 36) and F4 (lines 73-74), decided in round 1 (Prompt audit 2026-09-27: facts refuter exceptions, hash prose, split-table log check #443):
      • F1: role dispatch covers "each new or ad-hoc workflow agent() stage", and the vendored saved scripts keep their reviewed routing;
      • F4: "Its 2021 control segment has been inspected, so no experiment may present it as a fresh untouched holdout." This rewords a trading-section prohibition without weakening it.
    • X5b (line 3): the top rule is the operator's 2026-09-28 paragraph, byte for byte (lane-a/round1/user-level-top-rule.txt, on main). Lane A: GPT-6 agree (0.99), Claude agree (0.85).
    • S1 (line 37): each session builds a PR description from .github/pull_request_template.md. The required sota-sources check fails a description without a non-empty SOTA sources section. Lane A: agree (0.97) and agree (0.80).
  • CLAUDE.md lines 3-8 (X9): the GPT-6 lane's text replaces the line that told the model to use /mcp and /context, which it cannot run.
    • Round 3 ran the comparison that round 2 named: two cases, K5 and K6, in 30 headless runs.
    • Then all four blind judgments chose this text: GPT-6 and Claude, each in both orders.
    • main's ## Compact Instructions section (Claude Code practice for Opus 5.5: 2026-09-27 review resolved, host read-back, weekly reference freshness lane #453) follows unchanged.
  • docs/harness-defaults.md:
    • line 7 restates the new top-rule heading, a co-change with X5b;
    • four dated anti-pattern rows: round 2's three, and AN-13's headless run under the default print-mode wait.
  • The decision record docs/decisions/2026-09-27-prompt-audit-resolution.md gains two addenda:
    • X9's second round;
    • round 3: X9, lane A, AN-13, and the dashboard co-change.
  • Evidence under evidence/artifacts/prompt-audit-20260927/: x9-round2/ (77 files), x9-round3/ (65), an13/ (3), review-444/ (10), and five receipts in controls/.
  • The Codex lane's host receipt for Top rule templates: carry the operator's 2026-09-28 top rule (X5a, X5c) #458's template change (evidence/artifacts/codex-worker-lane-host-20260928/): the apply on this workstation, and a recheck with exit codes.
  • The grand dashboard's read bound, now on main through Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446.
    • This PR's registrations took the manifest past the 2,000,000-byte bound the dashboard applied to every source.
    • Its dashboard commit gave the manifest its own 8,000,000-byte bound.
    • Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446 crossed the same bound first and landed the same progress.py and test hunks in 3058b237 (two-file patch-id 8f693abd4c601190).
    • After the rebase, the dashboard commit keeps only the decision record's subsection and the two before/after receipts.
  • AN-13's M4 and M5 now point to their second-family lanes, which ran in AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460 (docs/decisions/2026-09-28-an13-m4-m5.md, merged). The M4 pointer also says that AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460's record leaves the skills trial's immediate-removal question open.
• Base commit: 8d8f79cd (AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460). Head: ed36bf61, eleven commits:
  • 787f3c6f: AGENTS.md F1 and F4;
  • d2afed0d: X9's second round;
  • f3bfd5cc: round 3 (X9 applied, S1 and X5b, AN-13's receipt);
  • 8a41f25e: the print-mode row, the line-7 co-change and the round-3 token receipts;
  • f08e290c: the dashboard bound's record subsection and receipts (its code is on main through Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446);
  • 033e679c: the review round's repairs, the audit's control, the Codex lane's host receipt and the review receipts;
  • b26eaa91: the first rebase note;
  • 52047445: the second rebase's notes: Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446 carried the dashboard bound, and review-444/README.md now names the reviewed commits by subject;
  • 26b81b8c: AN-13's M4 and M5 point to their 2026-09-28 lanes (AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460);
  • 95aecdb6: the M4 pointer says the trial's immediate-removal question is open, as AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460's merged record does;
  • ed36bf61: registration only.
• Rebases.
  • The first six commits were written on fb14dedf.
    • The branch was rebased onto 9f8db582 (fix(evidence): verify scheduled memory reports and scope digest exceptions #400) before its first push.
    • It was rebased again onto 3058b237 after Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446 merged, and a third time onto 8d8f79cd after AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460 merged.
    • Each time, main's manifest change conflicted with the registration. GitHub does not run pull_request workflows while a pull request has a merge conflict (events that trigger workflows), and validate.yml has no other trigger for branches.
  • fix(evidence): verify scheduled memory reports and scope digest exceptions #400's two anti-pattern rows are kept byte for byte.
  • Each registration was redone on main's manifest.
  • git range-diff 9f8db582..2c88180c 3058b237..ca55beb7 shows every patch unchanged except the dashboard commit, which lost the two code hunks Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446 had landed:

    1:  c4f8ba3c = 1:  e1d653d5 AGENTS.md: scope role dispatch to new or ad-hoc stages; current-state holdout wording
    2:  ce373a79 = 2:  3ff141bc X9 second round: blind two-family re-judgment splits again; CLAUDE.md unchanged
    3:  54f74628 = 3:  45edc9af Prompt audit round 3: X9 applied to CLAUDE.md; lane A's S1 and X5b in AGENTS.md; AN-13 receipt
    4:  2f41d8cf = 4:  98c03f21 Prompt audit: print-mode anti-pattern row (AN-13), line-7 co-change with X5b, round-3 token receipts
    5:  499a661e ! 5:  e738c9a4 Grand dashboard: the evidence manifest gets its own read bound
    6:  694473d6 = 6:  f3c07612 Prompt audit round 3: review repairs, the audit's control, the Codex lane host receipt
    7:  2c88180c = 7:  ca55beb7 Prompt audit: map the reviewed commits to the rebased ones
    

• Third rebase, onto 8d8f79cd. git range-diff 3058b237..51be81af 8d8f79cd..26b81b8c shows eight of the nine patches identical. The other, X9's second round, changed only in a context line, which is now AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460's anti-pattern row above this PR's rows. AGENTS.md and CLAUDE.md are byte-identical to eb12630a's (git diff eb12630a ed36bf61 -- AGENTS.md CLAUDE.md is empty).
• Lane: lane:shared, because it changes what the shared hot files AGENTS.md and CLAUDE.md say (docs/lanes.md).
  • The trading lane acknowledged it. native-agent-stack-84, the session running the trading lane, reviewed the F4 hunk at eb12630a and posted no objection (comment of 2026-09-28T16:04:57Z).
  • The third rebase left AGENTS.md byte-identical, so that acknowledgement covers this head's F4 hunk.
• Paths:
  • AGENTS.md and CLAUDE.md;
  • docs/decisions/2026-09-27-prompt-audit-resolution.md and docs/harness-defaults.md;
  • evidence/artifacts/prompt-audit-20260927/: x9-round2/, x9-round3/, an13/, controls/, review-444/;
  • evidence/artifacts/codex-worker-lane-host-20260928/: the Codex lane's host receipt for Top rule templates: carry the operator's 2026-09-28 top rule (X5a, X5c) #458's template change;
  • manifests/evidence.json: registration only, in the last commit.
  • No generated report changed.
• Size: o200k counts (gpt-tokenizer 3.4.0, tools/token-report/token_manifest.py count_files), from fb14dedf to this PR (controls/token-counts-r3-base.txt, -after.txt). 9f8db582 has the same AGENTS.md and CLAUDE.md as fb14dedf:
  • AGENTS.md: 2,374 to 2,484 tokens (+110); 103 lines, unchanged;
  • CLAUDE.md: 96 to 162 tokens (+66); 11 to 15 lines.

SOTA sources

• F1: the repository's dispatch rule source, examples/claude-native/workflows/README.md (its opening paragraph and the "Dispatch by role" bullet). Also Prompting Claude Sonnet 5: "If you need Claude to apply an instruction broadly, state the scope explicitly".
• F4: the same guide on explicit scope, and OpenAI: Using GPT-6, which says GPT-6 Astra is more sensitive to instructions in files such as AGENTS.md.
• X9, round 3 (x9-round3/sources-r3.json, Q1 and Q2, retrieved 2026-09-28):
  • The installed client, Claude Code 2.1.283:
    • each run's init event;
    • four uncounted probes (x9-round3/probe/facts.json);
    • its changelog at anthropics/claude-code@7779afb1.
  • Claude Code's hooks, headless and MCP pages. mcp.md L363 says stdio servers are not reconnected automatically; probe p2 saw 2.1.283 restart one.
  • The MCP specification's tools page (2026-07-28).
  • The K6 fixture server is built on the official MCP Python SDK v2.2.0 (modelcontextprotocol/python-sdk@9972c21a).
  • Harnesses:
    • promptfoo 0.123.1, running claude -p, for the comparison;
    • the packaged runner tools/sota-convergence/landscape-sweep/codex_call.sh, for the GPT-6 judges: codex-cli 0.157.1, gpt-6-astra, effort max, read-only;
    • the blind-adjudicator agent, for the Claude judges.
• X9, round 2: x9-round2/sources-round2.json (38 dated sources) and Claude Code skills: "Custom commands have been merged into skills."
• S1 (sources-r3.json Q3):
  • .github/pull_request_template.md, and the sota-sources check's heading rule (.github/workflows/validate.yml:544-545 on this base);
  • the live ruleset's required checks (lane-a/round1/facts-s1.json);
  • cli/cli v2.101.0 (pkg/cmd/pr/create/create.go L345-350, L619-641) and gh pr create --help: --body, --body-file and --fill skip the repository's PR template;
  • two maintained AGENTS.md files with PR instructions:
    • NVIDIA/warp@e477c3ba L26-28 says to use the repository's PR template;
    • WordPress/gutenberg@ff0659d6 L99-102 names CI enforcement.
• X5b: the operator's user-level paragraph of 2026-09-28 (lane-a/round1/user-level-top-rule.txt), and lane A's frozen packet and both lanes' returns (lane-a/round1/).
• AN-13:
  • Claude Code's environment variables page, for CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS, cited in the new log row and in an13/README.md.
  • Its commands and headless pages (sources-r3.json Q4): /doctor [prompt-audit [path]] is a skill, and user-invoked skills work in -p mode.
• Dashboard bound: no upstream component changes. The fix keeps the reader's own bounded-read design and gives its one growing source, the evidence manifest, a bound of its own (the record's last subsection).
• Procedure: docs/convergence-architecture.md. A text is applied only when both families choose it in both orders. The GPT-6 lanes and judges ran through the packaged runner.
• Withholding: the records rule (AGENTS.md, "Hosts, credentials and records"): evidence carries "no raw conversations, tokens, personal paths or machine-specific active client configuration".

Evidence-class table

Claim	Evidence class	Command / receipt
F1 and F4 were decided by both families	source_review plus retained model judgments	#443: round1/, adjudication/, decisions.json
S1 and X5b converged in lane A	source_review plus retained model judgments (one GPT-6 runner job, one Claude evidence-reviewer)	lane-a/round1/, on main
AGENTS.md:3 is the operator's paragraph byte for byte	local check	cmp below
X9's second round split	retained model judgments	x9-round2/attempt2/tally.json, audit.json
K5 and K6 results per text	local_integration: 30 headless runs of the installed client on Opus 5.5, with fixtures	x9-round3/analysis-r3.json, results-r3-compact.json
What the fixtures did in the client	local_integration: four uncounted probes	x9-round3/probe/facts.json
X9's round 3: four of four judgments for the GPT-6 lane's text, with a clean audit	retained model judgments: two GPT-6 runner jobs, two blind-adjudicator agents	x9-round3/judges/tally.json, audit.json
AN-13's two headless runs and each finding's disposition	local_integration: the installed client's own result records; the report text is not published	an13/an13-runs.json, an13/README.md
Token deltas	exact count of the files, not billed usage	controls/token-counts-r3-*.txt
The judges' audit can fire: clean copies reproduce its verdicts, and one planted access voids each judgment	local check on private copies of the judges' inputs	x9-round3/judges/audit-control.json, from audit_control_r3.py
The Codex lane's managed block is in place on this workstation after #458	local_integration on the real Codex home: dry run and static proof, no model call	codex-worker-lane-host-20260928/
The dashboard refused this PR's manifest, and the fix (now on main through #446) restores it without widening other sources	local_integration: unit tests, seen failing first	controls/dashboard-bound-before.txt, -after.txt
CI-equivalent	local	See below

Local commands run

# head ed36bf61, on 8d8f79cd; TMPDIR outside /tmp; the validate job's local steps in CI order
# (ci_local.sh without unittest; skipped here: the pip, actionlint and uv installs and their online audits)
workflows-sha256       exit=0
workflows-contracts    exit=0   (SUMMARY passed=254 failed=0 total=254; SUMMARY passed=74 failed=0 total=74)
zizmor                 exit=0
validate               exit=0   ({"components": 69, "hashed_files": 7894, "profiles": 4, "receipts": 163, "status": "passed"})
host-receipts          exit=0
release-due            exit=0
validate-catalogs      exit=0
validate-foundation    exit=0
landscape              exit=0
audit-reports          exit=0
validate-convergence   exit=0
build-verdicts         exit=0
component-matrix       exit=0
grand-list             exit=0
verdict-flip           exit=0
gap-crosswalk          exit=0
gap-wave-ledger        exit=0
build-ecosystem        exit=0
skill-contract         exit=0

# head ed36bf61: the modules that read the changed log, the dashboard and the installed profile
$ python3 -B -m unittest tests.test_adoption_docs_consistency tests.test_grand_dashboard tests.test_install_claude_profile
exit 0: Ran 112 tests; OK (skipped=4)

# full suite on 42ba6656, the head before the first rebase (fb14dedf plus the same six patches and a registration);
# CI's validate job ran it on 2edd5eaf's merge commit (Ran 6885 tests, OK (skipped=710)) and runs it on this head's
$ python3 -B -m unittest
exit 1: Ran 6882 tests in 717.410s; 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 dashboard bound, on this PR's manifest (2,017,827 bytes when these ran), before the rebases; the same code is now on main
$ python3 -B -m unittest tests.test_grand_dashboard   # main's progress.py
exit 1: Ran 17 tests; FAILED (failures=1, errors=8), each "ValueError: source exceeds size limit"
$ python3 -B -m unittest tests.test_grand_dashboard   # this PR's progress.py
exit 0: Ran 17 tests; OK

# in evidence/artifacts/prompt-audit-20260927/x9-round3/
$ sha256sum -c prereg-r3.sha256
exit 0: prereg-r3.json: OK
$ python3 -B grader_selftest.py
exit 0: 33 of 33 classified as labelled
$ python3 -B check_orders_r3.py judges/packets judges/mapping.json
exit 0: same evidence in both orders: True | other packet keys identical: True
$ python3 -B check_orders_r3.frozen.py judges/packets judges/mapping.json   # the frozen copy, kept to show its flaw
exit 1: same evidence in both orders: False
$ python3 -B exposure_scan_dir.py .
exit 1 by design: 16 files; every match is a name or pattern text (the package README's "Settings content")
$ python3 -B audit_control_r3.py <judges> <claude-AB> <claude-BA> <work> judges/audit.json judges/audit-control.json
exit 0: clean matches the published audit: True | every planted judgment void: True   # inputs are private copies

# evidence/artifacts/codex-worker-lane-host-20260928/, from the repository root at 0779e91a (tools/adoption/ as on fb14dedf)
$ python3 codex_lane_receipt.py <repo> <private> <out>
dry-run exit 0 (13:30:57Z to 13:31:14Z): "AGENTS.md: block already in place"; rehearsal passed
prove exit 0 (13:31:14Z to 13:31:22Z): result: PASS (7 pass, 0 fail)

# repository root
$ sed -n 3p AGENTS.md > line3.txt && cmp line3.txt evidence/artifacts/prompt-audit-20260927/lane-a/round1/user-level-top-rule.txt
exit 0 (736 bytes each)
$ python3 -B evidence/artifacts/prompt-audit-20260927/lane-a/verify_lane_a.py
exit 0: all checks passed

The two unittest failures come from this host, not the change. Both fail the same way when run alone on a clean worktree of main (lane-a/review/main-host-failures.txt, on main). Before the dashboard commit, the same suite also failed the eight dashboard tests: Ran 6881 tests, FAILED (failures=2, errors=8, skipped=765).

Reviews

• One review round on 0779e91a, before the first push (evidence/artifacts/prompt-audit-20260927/review-444/):
  • GPT-6: codex exec -s read-only -m gpt-6-astra, effort max; input 1,884,466 (1,797,760 cached), output 15,355. Verdict changes-needed, with one should-fix finding. It also found these consistent with their sources: the texts byte for byte, the token receipts (it re-ran gpt-tokenizer 3.4.0), the X9 retally, the registration, and the dashboard test against main's reader.
  • Claude: one evidence-reviewer (Opus 5.5; 77 API calls; cache read 11,985,562, output 102,328). Verdict changes-needed, with five should-fix findings and three nits.
  • Every finding concerned the decision record or the evidence text, not the instruction edits or the dashboard code.
• One repair round, f3c07612 (694473d6 and 6a3fb281 before the rebases), fixed every finding. review-444/README.md gives each finding's disposition. The round added two pieces of evidence:
  • the audit's discriminating control;
  • the Codex lane's host receipt, with exit codes and times.
• Not changed: GPT-6's note on trailing spaces in the retained judge prompts. The prompts are kept as the judges received them.
• No second round. The repairs had no independent review.
• Found before the first push:
  • the pre-push scan found the host user name once, in round 3's exposure_scan_dir.py;
  • the packager now writes it as <user>, as lane A's does;
  • the fix is folded into round 3's commit (45edc9af; 54f74628 before the second rebase), so no pushed commit carries the name. The same scan on the pushed head finds nothing: no home path, host user name, email, session scratch path, secret-shaped string or subagent id.

Decision record

docs/decisions/2026-09-27-prompt-audit-resolution.md, added by #443. This PR adds X9's second-round addendum and the round-3 addendum. X5a and X5c, lane A's other two items, landed in #458 with their own record, docs/decisions/2026-09-28-top-rule-templates.md.

Residuals

• The trading lane's acknowledgement is in (see Lane). If main moves before merge, rebase onto it and re-register manifests/evidence.json from main's copy.
• The dashboard bound landed through Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446 (3058b237). The host's live dashboard checkout gets it when that checkout is updated to 3058b237 or later.
• X9: interactive sessions are untested.
  • They are this decision's overturn condition, and only the user can run them.
  • The runs were headless, on Opus 5.5, under bypassPermissions, with ba1700ad's files.
• X8 is permission text: "after the current foundation work" is undefined in eight places. Only the user can define or close it.
• AN-13, for the user:
  • M1 and M2: the top rule's broad trigger, in the user-level file that AGENTS.md:3 copies by X5b;
  • M3: a pinned model version in the user-level file.
• AN-13, project-side: AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460 (merged) ran the second-family lanes for M4 and M5.
  • M4: no edit to the vendored text. AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460's record leaves the skills trial's immediate-removal question open: a two-family finding that the skill's instructions conflict with CLAUDE.md or AGENTS.md removes it under the trial's rule. Otherwise the preload question stays with the trial's 2026-10-25 comparison.
  • M5: split, so the lines are unchanged and both texts are recorded.
• The Codex lane on other hosts. Each host that installed the lane needs its own dry run and apply after Top rule templates: carry the operator's 2026-09-28 top rule (X5a, X5c) #458. It is done on this workstation (codex-worker-lane-host-20260928/).
• Earlier commits. Round 2's a28f6a91 and its ancestors stay reachable by SHA and carry K4's answer heads, which quote client configuration. Removing them from GitHub needs GitHub Support, and that is the maintainer's decision.

Host evidence

Not applicable: no file under evidence/hosts/ changed.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (none changed).
• New/changed workflows declare top-level permissions: contents: read (none changed).
• No secrets are printed, logged or committed; no new required secret.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code

[seathatflowsinourveins]

CI on the rebased head 2edd5eaf

• Validate run 36431757420 concluded success. Its validate job ran the full suite: Ran 6885 tests in 696.282s, OK (skipped=710).
• Every other check passed. validate-macos was still pending when this was posted.
• Run 36431564982 was cancelled when the title and body edit fired edited (cancel-in-progress).
• The local full suite in the body ran on 42ba6656, before the rebase. It showed only the two known host failures, and CI shows neither.

Merge-time steps (docs/lanes.md), after the trading lane's acknowledgement:

1. Replay the content commits on the new main without the old registration. In a fresh worktree:

   git checkout -B claude/prompt-audit-instructions-20260927 origin/main
   git cherry-pick 9f8db582..2c88180c

   If Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446 has merged first, main already carries the bound, so skip 499a661e (the dashboard bound, whole-commit stable patch-id 4fc1f751d522fa77). Run git cherry-pick --skip when the cherry-pick stops on it, whether it stops as empty or as a conflict in observability/grand-dashboard/progress.py or tests/test_grand_dashboard.py. Then check the other commits with git range-diff 9f8db582..2c88180c origin/main..HEAD.
2. Register again from main's manifest:

   git restore --source=origin/main --staged --worktree manifests/evidence.json

   Then register every file in git diff --name-only --diff-filter=AM origin/main...HEAD except the manifest with scripts/host_receipts.register_file. Run scripts/component_matrix.py --write, scripts/new_host_grand_list.py --write and scripts/validate.py, and commit the registration last.
3. Push with --force-with-lease. Then confirm mergeable: MERGEABLE and a validate run on the new head (gh run list --commit <sha>). GitHub runs no pull_request workflow while a pull request has a merge conflict.

[seathatflowsinourveins]

Rebased onto 3058b237 (#446). The head is now eb12630a.

Correction to my earlier merge-time steps: do not skip the dashboard commit.

• Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446 took only the dashboard commit's progress.py and test hunks. Their two-file patch-id is 8f693abd4c601190 on both sides.
• The commit also carries the decision record's dashboard subsection and the two controls/dashboard-bound-*.txt receipts, so skipping it would drop them.
• Replaying it onto a main that already has those hunks applies cleanly and keeps the rest. That is what this rebase did.
• git range-diff 9f8db582..2c88180c 3058b237..ca55beb7 shows 499a661e ! e738c9a4, and every other patch as =.

This push adds:

• 68dcc3cc: the record says Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446 carried the bound, and review-444/README.md names the reviewed commits by subject, so later rebases don't stale it;
• 51be81af: AN-13's M4 and M5 point to AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460, where their second-family lanes ran;
• eb12630a: registration on 3058b237's manifest, 172 files. validate.py passes.

Merge-time steps now, after the trading lane's acknowledgement (docs/lanes.md):

1. In a fresh worktree, replay every content commit, which is everything except the last registration commit:

   git checkout -B claude/prompt-audit-instructions-20260927 origin/main
   git cherry-pick 3058b237..51be81af

   Skip nothing. If AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460 has merged first, its only overlap is manifests/evidence.json, which step 2 rebuilds.
2. git restore --source=origin/main --staged --worktree manifests/evidence.json. Then register every file in git diff --name-only --diff-filter=AM origin/main...HEAD except the manifest, run component_matrix.py --write, new_host_grand_list.py --write and validate.py, and commit the registration last.
3. Push with --force-with-lease. Then confirm MERGEABLE and a validate run on the new head.

Local results on eb12630a:

• The validate job's 19 local steps all exit 0.
• tests.test_adoption_docs_consistency, tests.test_grand_dashboard and tests.test_install_claude_profile: 112 tests, OK.
• The pre-push scan finds nothing.

[seathatflowsinourveins]

Trading lane acknowledgement for F4 (docs/lanes.md shared-file rule). Reviewed at head eb12630a by native-agent-stack-84, the session running the trading lane today.

F4 is the only trading-semantic hunk: AGENTS.md, in the simulation-research paragraph about the 2021 control segment.

• Before: "Its 2021 control segment is now inspected; future experiments must not call it a fresh untouched holdout."
• After: "Its 2021 control segment has been inspected, so no experiment may present it as a fresh untouched holdout."

Verdict: acknowledged, no objection. The prohibition is not weakened. If anything it is slightly stronger:

• The scope widens from "future experiments" to "no experiment".
• The act widens from "call it" to "present it as", which also covers framing that stops short of naming it.
• "No experiment may…" is an unambiguous prohibition.
• The inspected fact is unchanged, and "so" now states the reason for the rule.

I reviewed only this hunk. The rest of #444 is foundation instruction text and hash re-registration, outside the trading lane.

[seathatflowsinourveins]

Rebased onto 8d8f79cd (#460). The head is now ed36bf61.

• git range-diff 3058b237..51be81af 8d8f79cd..26b81b8c: eight of the nine patches are identical. The other, X9's second round, changed only in a context line, because AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460's anti-pattern row now sits above this PR's rows.
• The one new content commit is 95aecdb6. The AN-13 M4 pointer now says that AN-13 M4/M5: two-family lanes; M4 kept, M5 split after blind adjudication; judge-dispatch anti-pattern row #460's record leaves the skills trial's immediate-removal question open.
• AGENTS.md and CLAUDE.md are byte-identical to eb12630a (git diff eb12630a ed36bf61 -- AGENTS.md CLAUDE.md is empty). The trading lane's F4 acknowledgement at eb12630a therefore covers this head.
• The registration was redone on main's manifest: 172 files, and validate.py passed with hashed_files 7894.
• This replaces my earlier merge-time steps. If main moves again:
  1. Cherry-pick 8d8f79cd..95aecdb6 onto the new main, skipping nothing.
  2. Re-register from main's manifest as the last commit.