Lane prompt Rule 2: status, license and popularity are not evidence of fit (D1 merit rule)

[seathatflowsinourveins]

Summary

This PR implements runbook decision D1, as the A2 rules-only PR. It adds one sentence to Rule 2 of the shared lane prompt tools/sota-convergence/lane-prompt.md. The sentence is generic and names no repository or product:

Adoption, incumbency, installation, retained-control status, receipt count, packet position, license, stars and popularity are not evidence of fit; choose among adopted candidates on what their evidence shows was run.

• Scope of the sentence.
  • The runbook's D1 example lists adoption, installation, receipt count and packet position. The sentence adds the two terms of the user's merit directive that the example leaves out: incumbency and retained-control status.
  • It also names license, stars and popularity, following two later user directives of 2026-09-26 (quoted verbatim, typos included):
    • "license are not applibalbe for our env,as we are priviate experimental env not company"
    • "keep it so the future session can avoid gates unnessary , focuing on the quality and sotaness of the repos itself compare to its candidates in that field"
  • Adoption stays what Rule 2's first sentence makes it, the gate for entering the winner set. The new sentence says adoption is not merit.
• Still five rules. The prompt keeps exactly five numbered rules. lane_packets.load_rules parses them into every packet's rules field, so every packet built from this head carries the sentence. That includes the packets read by the Claude refuters and the blind adjudication judges.
• Prompt hash. The prompt's prompt_sha256 changes from 9b8a8364f13ff3fcb9fc690198c4ca8849c29cc12b81e8526121a46114517f54 to d3ef2cc0c06d33d6a4a058f5e74a1cf5be8a3760cc2b0b5572598f0e0e4288e4.
• Registry appends. tools/sota-convergence/lane-provenance.json only gains entries. Every existing entry keeps its bytes, because those entries bind the earlier waves.
  • Two claude entries, one for .claude/workflows/layer-verdict-lane.js and one for the vendored examples/claude-native/workflows/layer-verdict-lane.js. Each copies the 2026-09-24 pair: workflow fd77b749…, role a87a9883…, transcript audit 5ed7a1b5… and vendored_path. Only the prompt hash (the new one) and registered_at (2026-09-26) differ.
  • One codex entry, pairing the current codex_lane.py (a4bd8f47…) with the new prompt hash.
  • No adjudication entry. That entry's prompt_sha256 hashes adjudication-prompt.md (f82b527d…, registered by GPT-6 live web search, Claude WebSearch session budget, and GPT-6/Claude family names in the blind-adjudication leak checks #352), and this PR changes neither that file nor codex_lane.py.
  • After the rebase, each of main's lists is an exact prefix of this head's list, GPT-6 live web search, Claude WebSearch session budget, and GPT-6/Claude family names in the blind-adjudication leak checks #352's adjudication append included. git diff origin/main removes no line of the file.
• Tests.
  • tests/test_lane_packets.py gains LanePromptMeritRuleTests:
    • Rule 2 must carry the sentence.
    • The Rules section must number exactly 1-5. This test counts the rules itself, separately from load_rules, which only counts them.
    • The existing packet test now asserts that every packet's rules[1] carries the sentence.
  • tests/test_verdict_lane_vendoring.py gains test_the_current_claude_lane_key_is_registered_under_both_workflow_paths.
    • It checks the full record-time Claude key, scripts/landscape.py LANE_PROVENANCE_KEYS["claude"], prompt hash included, under both workflow paths.
    • Until now only the Codex prompt hash was tested, so a missing Claude append surfaced only when a verdict was recorded.
  • test_registry_entries_are_well_formed_and_unique now keys each lane on LANE_PROVENANCE_KEYS.
    • That is the key record time matches on.
    • The old Claude tuple left out prompt_sha256, so this prompt-only append collided with the 2026-09-24 pair and failed the test (10 != 8).
    • The same widening was made for an audit-only append in round 11 (RR11-2).
    • The test now also checks the format of the Claude prompt_sha256.
• Evidence registration. GPT-6 live web search, Claude WebSearch session budget, and GPT-6/Claude family names in the blind-adjudication leak checks #352 (59dfe49f) registered tools/sota-convergence/lane-provenance.json in manifests/evidence.json after this branch was first cut. This PR follows the hot-file protocol in docs/lanes.md:
  • It takes main's manifest and re-registers that one file with host_receipts.register_file. Only the entry's sha256 and bytes change.
  • Before the re-registration, scripts/validate.py failed on the file (SHA-256 mismatch, byte count mismatch). After it, the validator passes.
  • scripts/component_matrix.py --write and scripts/new_host_grand_list.py --write ran and left their reports unchanged.
  • The other three changed paths are not registered in main's manifest.

Scope

• Base commit: 59dfe49f, the merge of GPT-6 live web search, Claude WebSearch session budget, and GPT-6/Claude family names in the blind-adjudication leak checks #352 and origin/main at the rebase.
• Lane: lane:shared.
  • The prompt judges both catalogs' layers.
  • lane-provenance.json is a verdict trust path (scripts/verdict_review_gate.py TRUST_PATHS).
  • manifests/evidence.json is a shared hot file. This PR only re-registers one hash in it.
• Lane consolidation: on 2026-09-26 the user shut down the other lane sessions and assigned this work to the workstation coordinator ("no other peers live, i shut them with the token save repos install into our native workflow and github workflow , harness update and sota foundation of our session as main task"); this quote stands in for the trading lane's ack.
• Paths touched:
  • tools/sota-convergence/lane-prompt.md
  • tools/sota-convergence/lane-provenance.json
  • tests/test_lane_packets.py
  • tests/test_verdict_lane_vendoring.py
  • manifests/evidence.json (one re-registered entry)
• No wave files, verdict rows, sealed artifacts or other evidence registrations change. verdict-review-gate treats a trust-path change without a verdict change as rules-only.

Ordering (runbook A0/A2)

• Merge before B1, which is the gate of A2.
  • After this lands, a lane run started on the old prompt cannot be recorded. claude_lane.py binds the echoed prompt to this lane-prompt.md, and codex_lane.py reuses a return only when its provenance matches.
  • Under the A0 hold, nothing that touches lane-prompt.md or lane-provenance.json merges from B1 to B10. If this PR misses B1, it waits until after B10, and the wave runs on the old prompt without the sentence. The wave's own record must then disclose that its lane prompt did not state the merit rule.
• GPT-6 live web search, Claude WebSearch session budget, and GPT-6/Claude family names in the blind-adjudication leak checks #352 has merged as 59dfe49f. This branch is rebased on it, so its adjudication append comes before this PR's entries.
• Refuse env-style shell launchers; keep unix and local URL schemes checked (#206 review lows) [hold: after the wave] #216 is held until after the wave. It changes codex_lane.py and lane-provenance.json. When it lands after this PR, its Codex append must pair its codex_lane.py hash with prompt d3ef2cc0…, not 9b8a8364…, and keep this PR's entries in place.

Disclosures and residuals

• Where the sentence does not appear.
  • The Claude lane's lens prompts in the vendored layer-verdict-lane.js do not state it. That file is hash-bound and vendored from agent-lab.
  • adjudication-prompt.md does not state it either.
  • The refuters and judges get the sentence only through the packet's rules field.
• The rule instructs; it does not blind.
  • Exported evidence prose is not rewritten, so a lane can still meet control-arm labels in comparison receipts.
  • The sentence tells a lane not to weigh stars, license or popularity; it does not remove such fields where the evidence carries them. For example, D2 makes observed.stars and pushed_at visible in newcomers' source reviews.
• Verdict lanes only. The sentence governs how verdict lanes judge fit. This PR changes no license field or check anywhere else in the repository.
• L3 rewording not bundled. The README's disclosed L3 rewording (overturn_when citing tests/ paths) is not part of this prompt revision.
• Effect not measured. No lane has run on the new prompt, so the effect on verdicts is unmeasured until the B-phase wave. The tests below are local checks and synthetic fixtures only.
• Review status. GPT-6 reviewed the pre-amendment head 476d00c3 and found no defects. The amended wording, the new hash, the rebase and the evidence re-registration came after that review and have had no independent review (see Review).

Evidence-class table

Claim	Evidence class	Command / receipt
The prompt parses to five rules, with the sentence in Rule 2 and in every packet's rules[1]	synthetic (fixture)	python -m unittest tests.test_lane_packets
The registry covers the current Claude key (prompt included) under both workflow paths, and the current Codex key	local_integration (checkout check)	python -m unittest tests.test_verdict_lane_vendoring
Main's registry lists are exact prefixes of this head's lists, and the extra entries are the reviewed appends with only prompt_sha256 changed	local_integration; not retained (a scratch comparison script, not committed)	compare git show origin/main:tools/sota-convergence/lane-provenance.json with this head, list by list
Each mutation below fails the named new or widened tests	synthetic; not retained (a one-off in-place mutation run with files restored byte for byte; its script was not committed, so this row is a statement of the author's run, not a receipt)	rerun it by applying any listed mutation and running the named test
Lanes weigh evidence over status, license and popularity on the new prompt	not established	the B-phase verdict wave

Mutation run on the amended tree (test names shortened; merit = test_rule_2_states_the_merit_sentence, count = test_the_rules_section_numbers_exactly_five_rules, packet = test_rules_are_the_five_lane_prompt_rules, claude_key / codex_key = the two registry-coverage tests, unique = test_registry_entries_are_well_formed_and_unique):

baseline {'merit': 'pass', 'count': 'pass', 'packet': 'pass', 'claude_key': 'pass', 'codex_key': 'pass', 'unique': 'pass'}
drop sentence {'merit': 'FAIL', 'packet': 'FAIL', 'count': 'pass'}
move sentence to rule 3 {'merit': 'FAIL', 'packet': 'FAIL', 'count': 'pass'}
sixth rule {'count': 'FAIL', 'merit': 'FAIL'}
renumber 5->6 {'count': 'FAIL'}
reviewed 476d00c3 wording {'merit': 'FAIL', 'packet': 'FAIL', 'count': 'pass'}
no license {'merit': 'FAIL', 'packet': 'FAIL', 'count': 'pass'}
no stars {'merit': 'FAIL', 'packet': 'FAIL', 'count': 'pass'}
no popularity {'merit': 'FAIL', 'packet': 'FAIL', 'count': 'pass'}
no claude appends {'claude_key': 'FAIL', 'codex_key': 'pass', 'unique': 'pass'}
only vendored-path claude append {'claude_key': 'FAIL'}
no codex append {'codex_key': 'FAIL', 'claude_key': 'pass'}
duplicate claude append {'unique': 'FAIL'}
appends keep the pre-amendment hash {'claude_key': 'FAIL', 'codex_key': 'FAIL', 'unique': 'pass'}
restored True

Decision record

No dated decision record is committed. The decision is D1 in the coordinator session's verdict-wave runbook of 2026-09-26, amended the same day by the two user directives quoted above.

• Adopted: the sentence above, with license, stars and popularity.
• Alternatives considered:
  • The D1 example sentence verbatim. It leaves out two of the merit directive's named terms.
  • The reviewed 476d00c3 wording, without license, stars and popularity. The two later directives superseded it.
  • Declining D1. The wave's record would then disclose that the prompt does not state the rule.
• What would overturn it:
  • The B-phase wave's lane returns keep choosing the incumbent on layers where a non-incumbent adopted candidate has the stronger executed evidence. That would show the sentence is not enough and that the lens and adjudication prompts need the rule as well.
  • For the license term: this environment stops being private and experimental (for example, company use or redistribution). License would then be a real constraint again, and "license" should leave the sentence.

Local commands run

These ran on this host with a CI-equivalent virtualenv (Python 3.13.15 with jsonschema, PyYAML and numpy). The tests, the mutation run and the validators ran on the working tree that was then committed unchanged as e0ec2d29. The gate and the full suite ran on that commit.

$ python -m unittest tests.test_verdict_lane_vendoring tests.test_lane_packets tests.test_codex_lane
Ran 207 tests ... OK                                               # rc=0, 0 skipped: the A2 gate
$ python -m unittest tests.test_claude_lane tests.test_record_verdicts
Ran 146 tests ... OK                                               # rc=0, 0 skipped: the other modules that read lane-prompt.md
$ python scripts/validate.py
{"components": 69, "hashed_files": 7109, "profiles": 4, "receipts": 159, "status": "passed"}
$ python scripts/evidence_manifest.py --check
{"files": 7109, "status": "passed"}
$ python tools/sota-convergence/build_verdicts.py --check --root .
{"status": "checked", "waves": ["20260922"], ...}
$ python scripts/landscape.py --root .
{..., "status": "passed"}
$ python scripts/verdict_review_gate.py --root . --base origin/main
verdict-review-gate: no verdict rows changed (base 59dfe49f2334)   # rc=0
$ python -m unittest discover -s tests -q
Ran 6350 tests in 573.370s
OK (skipped=665)                                                   # rc=0; git status --porcelain empty afterwards
$ sha256sum tools/sota-convergence/lane-prompt.md
d3ef2cc0c06d33d6a4a058f5e74a1cf5be8a3760cc2b0b5572598f0e0e4288e4   # the prompt_sha256 of all three appended entries

The pre-commit gitleaks hook reported no leaks.

SOTA sources

• The user's directives (2026-09-26). The sentence folds them into the prompt generically. The merit directive's named example stays out of the prompt.
  • Merit, relayed by the landscape-sweep session: "make sure every layer the winner repos earns it, no biases that assume ai-memory etc are the best but with true merit of its quality, SOTA-ness and evidence".
  • License: "license are not applibalbe for our env,as we are priviate experimental env not company".
  • Gates and comparison within a field: "keep it so the future session can avoid gates unnessary , focuing on the quality and sotaness of the repos itself compare to its candidates in that field".
• Verdict-wave runbook, coordinator session, 2026-09-26:
  • §1 D1: "Add the user's merit rule to lane-prompt.md … generically, with no names. Fold it into Rule 2 … append claude and codex entries to lane-provenance.json … keep exactly 5 numbered rules."
  • §2 A2: the rules-only PR and its test gate.
  • §2 A0: the hold on these paths from B1 to B10.
• Precedents for prompt and provenance appends (this repository):
  • PR-5 (7f138632, 2026-09-22) created lane-prompt.md, with five rules parsed by lane_packets.load_rules.
  • Verdict-pipeline integrity before the 20260923 re-record: per-wave CI, unbiased packets, two-family rule, survivorship, receipt-backed status #122 (97a56e22, 2026-09-23) created the append-only lane-provenance.json, keyed on (codex_lane_py_sha256, prompt_sha256).
  • Blind both layer-verdict lanes for the 2026-09-23 re-record #141 (2013d27e, 2026-09-23) folded the blind boundary into Rule 1 as a sentence so that "lane-prompt.md keeps five rules", and appended the Codex entry for prompt fa75f113….
  • Blind lanes round 2: sanitized export, receipt aliases, two-family adjudication #145 (ff2ac50a, 2026-09-24) revised the prompt and appended the Codex entry and both Claude entries for prompt 9b8a8364….
  • Round 11 RR11-2 in tests/test_verdict_lane_vendoring.py set the precedent for widening the uniqueness key.
• Hot-file protocol: docs/lanes.md, "Hot-file protocol". Take main's manifests/evidence.json, re-register the branch's files with host_receipts.register_file, then run the two --write generators and scripts/validate.py.
• Same directive, earlier wording: tools/sota-convergence/landscape-sweep/templates.json, the "Merit, not incumbency" paragraph of the 2026-09-26 sweep. That wording names products; this prompt stays generic.
• Repository evidence rule: the top rule in AGENTS.md ("Stars, installs and popularity guide discovery; they are not evidence") and docs/harness-defaults.md ("Agreement, recency, stars and extra tooling are not evidence").

Review

• Cross-family review of the pre-amendment head, 2026-09-26. GPT-6 (gpt-6-astra) ran through codex exec at reasoning effort max in a read-only sandbox, on 476d00c3. That head was based on b21a403c, and its sentence did not yet name license, stars or popularity (prompt 60a4115a…).
  • Its scope was the runbook's D1, A0 and A2 text, the draft body as an unverified author claim, and five D1 checks: blindness; exactly five rules that still parse; provenance appended, not rewritten, with matching prompt hashes for both lanes; sealed waves still verify; test discrimination.
  • It exited 0 after 101,226 tokens. Its transcript has no usage-limit markers.
  • Verdict, verbatim: NO DEFECTS.
  • It inspected lane_packets.load_rules and scripts/landscape.py LANE_PROVENANCE_KEYS with Serena find_symbol, and ran scoped checks through context-mode: "The focused tests, provenance hashes, sealed-wave check, publication validator, and review gate passed."
  • It stated its own limit: "Those checks won't establish how a model responds to the new rule." That is the "Effect not measured" disclosure above.
  • One mid-review suspicion, a diff that looked like a rollback of the grand-dashboard checkpoint, was resolved by the reviewer: it was dde28cc2 landing on main after the branch was cut.
  • Findings: 0 held, 0 fixed, 0 rejected.
• Changes after the review. 476d00c3 was rebased onto 59dfe49f and amended into e0ec2d29. git range-diff b21a403c..476d00c3 59dfe49f..e0ec2d29 shows the whole change against the reviewed bytes:
  • lane-prompt.md: the sentence's two lines, now naming license, stars and popularity.
  • lane-provenance.json: the three appended entries' prompt_sha256, from 60a4115a… to d3ef2cc0…. The appends change in no other byte. Main's entries, GPT-6 live web search, Claude WebSearch session budget, and GPT-6/Claude family names in the blind-adjudication leak checks #352's included, come before them.
  • tests/test_lane_packets.py: MERIT_SENTENCE and its comment.
  • manifests/evidence.json: the lane-provenance.json entry's sha256 and bytes.
  • The commit message.
  • tests/test_verdict_lane_vendoring.py does not change.
• Not re-reviewed. No independent review has looked at these post-review changes. This is a recorded residual; the 476d00c3 verdict does not carry over to them.
• Evidence class. The review is an independent cross-family read-only review of the pre-amendment head. The reruns are local checks on this host. Neither measures how a lane behaves on the new prompt.
• Open before merge.
  • The PR's own CI.
  • Merging before B1 (A2 gate, A0 hold).
  • An independent review of the post-review changes, if the coordinator requires one.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

[seathatflowsinourveins]

Amendment review and replay (coordinator, 2026-09-26):

• GPT-6 cross-family round on the amendment (476d00c3 → e0ec2d29): gpt-6-astra, max effort, read-only, live search, rc=0. Verdict NO DEFECTS. The reviewer ran the 16 targeted tests and repository validation, and both pass. Removing any one of license, stars or popularity fails an existing test. The provenance hashes match, and no prior entry changed.
• Replay onto main 4a32e646 (Native data (G2): token-tool savings into Loki and Grafana, one series per native scope #365 merged) as a3858f2f, under the hot-file protocol: main's manifests/evidence.json with the four changed files re-registered. Their content is byte-identical to e0ec2d29. validate.py FAILS=0; verdict-review-gate reports no verdict rows changed; 596 core tests OK; the lane modules (vendoring, packets, codex_lane, claude_lane, record_verdicts) ran 353 tests OK.