docs: close seven W6 foundation layers with preregistered comparisons

[seathatflowsinourveins]

Scope

Record SCHEDULED terminal closeouts for seven W6 foundation layers after #1007 landed. Each layer has a comparison id, exact preregistration digest, named readiness event and fixed decision rule, with every cumulative survivor/pending proposal disposition and a native run-pin assessment at c2dee17.

• Base commit: 8b1ef0c37e2f523aa84d2c75960540f45e74e814 (PR base; c2dee17 remains the historical run authority).
• Lane: lane:foundation.
• Owned paths: docs/decisions/w6-practice-closeout-20261011/ and their native manifests/evidence.json registrations.

SOTA sources

• #1008 layer-resolution protocol, squash 569bbc59db2b5f7c8a932357726efb7acca90b91, terminal-state section 69–122.

• #1007 native final-stack reader, c2dee173da1f792ffd9cce958dd2ca84f47e8c16: Sources.json and run_pin, reading catalogs/landscape/final-stack.json at the same landed revision.

• CC ruling-w6-wave-dispatch-20261011.md §§2–4, SHA256 a13204da949ba7f1b883332ea601790a2570ccd6066e9bb1bbc150b7b2e03304; Generate the canonical final-stack manifest (W4 PR 2) #1007 LANDED relay at 23:35:33Z.

• anthropics/claude-code-action, 2dca132ff0e0c4094ce6048b422c6915a071210b: README.md, package.json, CONTRIBUTING.md; source file hashes and documentation section refs retained in primary-source-refs.json.

• ataraxy-labs/sem, 225e7721c6ce7e220f0e60783af8cd74c1881ed9: README.md, package.json, CONTRIBUTING.md; source file hashes and documentation section refs retained in primary-source-refs.json.

• UKGovernmentBEIS/inspect_ai, a1dbea8db3cc85e04ff532760d0eaaeb64df6d3a: README.md, Makefile, pyproject.toml; source file hashes and documentation section refs retained in primary-source-refs.json.

• promptfoo/promptfoo, 34f74d34e140b5e17d23770dfb2340057b1936b8: README.md, package.json, CONTRIBUTING.md; source file hashes and documentation section refs retained in primary-source-refs.json.

• github/gh-aw, c16656ef90fa17dfe6e91fa8fb3d11f54632a74d: README.md, Makefile, CONTRIBUTING.md; source file hashes and documentation section refs retained in primary-source-refs.json.

• openai/codex-action, a8d8be6ee028bea4525bc722d16f36687b408c81: README.md, package.json; source file hashes and documentation section refs retained in primary-source-refs.json.

• reviewdog/reviewdog, cfb9ce1577e1e86f66bdfef6113d27340340923e: README.md, package.json; source file hashes and documentation section refs retained in primary-source-refs.json.

• mergifyio/mergify-cli, 6307ed10e8628c6fefe408e398ea48af16c94cfd: README.md; source file hashes and documentation section refs retained in primary-source-refs.json.

• coderabbitai/skills, 928c3b61f71b578b1bcfbacc289f1446244dd9dd: README.md, CONTRIBUTING.md; source file hashes and documentation section refs retained in primary-source-refs.json.

• OpenHands/OpenHands, c65ab872ccb2974ae0b1f5965813bb50c4199132 (historical source; resolver role RETIRED): README.md, package.json; source file hashes and documentation section refs retained in primary-source-refs.json.

• SWE-agent/SWE-agent, 3ea751c087f32b16e039a2233dd6eefecef325d5 (superseded by mini-swe-agent; not a Review or Landing arm): README.md, pyproject.toml, CONTRIBUTING.md; source file hashes and documentation section refs retained in primary-source-refs.json.

• dagucloud/dagu, 58fed633d58c1dd1319091fdb2c2f6158ecfa053: README.md, Makefile, CONTRIBUTING.md; source file hashes and documentation section refs retained in primary-source-refs.json.

• cli/cli, ec5b512045db67e5a2a4ff4a1b02660b2fb24390: README.md, Makefile; source file hashes and documentation section refs retained in primary-source-refs.json.

• ollama/ollama, eab97e9f92b9a25c2d52d2cc6c1b1c99bd9fae21: README.md:11,52,90,141; lazy fetch and optional locks disabled. Source availability supplies no model-run acceptance.

• PR-Agent local CLI, README:88-98 and :189-191; retained grand-catalog PENDING row and seed remain source observations.

• Deprecated openai/skills, README:2; named successor openai/plugins. Source review only; no successor adoption.

• mini-swe-agent, with retained a83fcae8 source; supersession citation only, routed to the workers closeout.

Evidence-class table

Claim	Evidence class	Command / receipt
Every layer has a bounded terminal decision with an immutable comparison protocol	source_review	terminal-states.json and comparison-preregistrations.json
Recorded run pins match the landed aggregate's native readers	local_integration	scripts.final_stack.Sources.json/run_pin; 18 winner joins
Every cumulative survivor and missing-vote-only pending proposal has a disposition	source_review	inputs-and-dispositions.json; 21 unique entries: 20 comparisons and one retired role
Input digests, comparison rules/triggers and native registry entries agree	local_integration	Re-frozen hash/rule/trigger/role/pointer check; evidence_manifest --check
Repository and publication checks pass	local_integration	validate.py, final_stack.py --check, git diff --check and gitleaks

Terminal states

Layer	State	Comparison id	Exact preregistration SHA256
git-github-automation	SCHEDULED	w6-git-github-automation-20261011	d7c8542d5897c6dd2f96822a370996d95a00473ec8b97f245765558c7f36a3b8
instructions-skills	SCHEDULED	w6-instructions-skills-20261011	f85304bb7bdfac05d361c746babe6a3a62a5186ea06ee72302f4f3039b724774
observation-inference	SCHEDULED	w6-observation-inference-20261011	22c642b8c5d331edf19c9e84b52e381fb85b4315f24f5d2d7f4f80e6f05f82cf
quality-evaluation	SCHEDULED	w6-quality-evaluation-20261011	47782b9a7a6320d89c0c7a794998af7e05682395839c7495323fed8966df1716
recovery-portability	SCHEDULED	w6-recovery-portability-20261011	cb43f16a0044a883354209743f55c024ee8b97d647d3d8b2d963b06597c36c4a
scheduling-supervision	SCHEDULED	w6-scheduling-supervision-20261011	adc039640a8a52599b153aa00cd30276d4a5bdeb0bf418909e808e35183cf13b
token-efficiency	SCHEDULED	w6-token-efficiency-20261011	16e0fad18981e28b07686ec5efc4fd524539f7436f7ad55a7ef8da022173c8fb

The readiness event for each is retention of its hash-bound named input packet with every source, case/oracle and owner-receipt input listed in the comparison file. The preregistered review and landing replays also requires #953 and #975 landing; both were OPEN at the original source check and have since landed at 3a8a846 and 86fa384. Sixteen retained candidate source mirrors remained missing in the observed source snapshot. When all inputs exist, the committed comparison runs through the existing W5 harness; it is not relaunched as research.

The landed run authority contains 18 historical winners for these layers, with 16 run pins, two absent skill pins and six verdict/run pin differences. This closeout does not join an upstream-owned install/test E2E at every pin or a complete adoption-window invoke record, so no FINAL winner is claimed. Existing matrix host_e2e factors retain their scopes; no zero-call inference is made.

Verdict, sweep and survivor inputs

Consumed references are frozen from 77185991ec9627e59473df5b4d7a11ca87b0461d: catalogs/landscape/foundation.json, wave 20260922 in catalogs/sota-convergence/layer-verdict-waves.json, its detailed verdict artifact, the latest covering landscape-sweep-20260929 ledger rows, manifest and retained returns, the earlier adjudications, and the 10-08 grand catalog/index. The terminal/run authority is the later c2dee17 landing. The 10-09 Pine sweep covers another layer.

Layer	Cumulative unresolved disposition
git-github-automation	claude-code-action's in-GitHub Claude OIDC role retired by #981; sem included in the structural-diff comparison. Review and Landing have separate eligible arms and preregistered W1/landing-incident replays.
quality-evaluation	inspect_ai and MTEB included; Promptfoo versus Inspect uses the same 60 independently labeled advice cases, with MTEB's embedding-evaluation role explicit.
instructions-skills	mattpocock/skills included in selected-file paired task trials.
token-efficiency	sqz, paritok-4b-v1 and agent-console included; mcp-compressor and klavis missing-vote-only proposals remain included.
observation-inference	openllmetry, ollama, nvidia_gpu_exporter and dcgm-exporter included; mlx-lm and llama-swap missing-vote-only proposals remain included.
scheduling-supervision	dagu, pg_durable, restate and conductor included in the scoped owned-recovery comparison.
recovery-portability	microsoft/wsl missing-vote-only proposal included; same-host/synthetic and destination-host boundaries remain explicit.

This is 16 substantive survivors plus five missing-vote-only proposals. Later omissions do not clear survivors, and missing votes do not refute them. Every latest substantive adjudication and source-recovery gap remains referenced in the record.

Local commands run

python3 scripts/final_stack.py --check
exit 0; landed aggregate checked; existing report-only flags retained

Native cross-artifact input/hash/rule/trigger and Sources.json/run_pin checks
exit 0; 7 SCHEDULED states, 7 exact preregistration digests,
12 frozen input digests, 21 dispositions, 18 run joins (2 absent, 6 drift)

python3 scripts/evidence_manifest.py --check
exit 0; 11753 files

python3 scripts/validate.py
exit 0; 70 components, 11753 hashed files, 4 profiles, 239 receipts

git diff --check 8b1ef0c3 e1c885ccf2da76e5f1a1ba78816a4e1738d1dd40
exit 0

gitleaks dir docs/decisions/w6-practice-closeout-20261011 --redact=100 --no-banner
exit 0; no leaks

No heavy E2E, mutation run or model replay ran. Heavy suites in Monday 08:00Z–20:00Z use the verified T2 idle wrapper; full suites also require SUITE-REQUEST/SUITE-GO. Cued PR work and the paper path retain precedence.

Decision record

docs/decisions/w6-practice-closeout-20261011/README.md, with machine-readable states, comparison protocols, survivor dispositions and native run assessment beside it. The CC explicitly closes HARNESS-WIF-FIX-PROPOSAL as superseded by #981; its historical bytes remain intact. The git layer also retains gh 2.102.0's lifecycle/run-inventory discrepancy and the Codex pin_moved receipts as required comparison/qualification inputs.

Host evidence

Source and local integrity checks only. No host receipt, platform status, canonical catalog, generated aggregate, live configuration or paper unit changes.

Follow-ups

• Astra fix1 P3-2: add focused native regression coverage for the corrected declarative role boundaries, retained conditions, rule/readiness copies and retirement records, with pinned mechanism references and fail-before evidence; the reviewer's structural checks are not committed runtime acceptance.

• Astra fix1 P3-3: retain the cited pinned public captures and complete vendor mechanism/source/install/test packets before scenario execution. Publish only safe attributed authority bytes; preserve all missing-input gates.

• Astra fix1 P3-5: normalize inherited hostname identifiers and their registrations in the owning hygiene change, preserving evidence bindings; no hostname is introduced by this patch or body.

• Astra fix1 P3-6: obtain genuine-worktree results for the named additional archive/Git/history-dependent checks through their owning verification lane. Keep these gaps separate from the supplied thirteen-module precompute; do not weaken guards or count unavailable results as passes.

• Keep the 16 bounded primary-source recovery inputs explicit; recover their known pins without new discovery.

• Reconcile gh 2.102.0's recorded run pin and native lifecycle evidence through the owning inventory/receipt paths; preserve the c2dee17 snapshot.

• Qualify every winning arm through upstream install/test and invoke evidence before a later FINAL decision.

• Feed accepted terminal closeouts into the native aggregate's terminal-state integration through its owning supported mechanism; its existing generated statuses keep their meanings.

• Preserve historic source/measurement boundaries and owner receipts during subsequent comparisons; no broad acceptance follows from this closeout.

• Retain the public source/case/correction packets and primary source captures for readers; preserve every conditional dependency until its input exists.

• Keep metric units, price tables, oracle tolerances and exact tie/all-rejected behavior fixed before execution; do not tune on arm results.

• Normalize the inherited hostname identifiers with their registrations in the owning hygiene change.

• Obtain genuine-worktree qualifications for the skill-usage and currency-due archive-environment failures; preserve the base/head failures as unqualified.

Checklist

• Non-empty SOTA sources section, exact source pins and recorded source boundaries.
• Seven terminal states, comparison ids/digests/readiness/rules and every unresolved disposition.
• Landed native final-stack reader and run-pin joins used.
• Native evidence registrations and publication scans pass.
• Peer worktrees, frozen history and docs: consolidate research convergence and close practice workstreams #1011's frozen head preserved.
• Astra and CC Claude closeout reads, and the applicable same-head wave gate, complete.

CC-adjudicated role correction

Previously reviewed head 243c2c26d7a1e94e1997127f51d0513fc5683e4f carries ONE fix commit after both first reads, on the registry-only carry of b9ee9e6 to main 8b1ef0c37e2f523aa84d2c75960540f45e74e814. The original non-registry row is unchanged. Review has five eligible arms; Landing has two and its preregistered merge-skew/thread/verdict/false-stop/time incidents. Neither resolver nor the reporting-only reviewdog is an arm; native merge queue awaits the org move.

The correction preserves each retained candidate protocol under the role it covers, including GPU metrics, model routing, usage attribution, MCP permission/approval/hook/invoke/instruction gates, Dagu 2.17.2 scheduler liveness, three-host Restate recovery, SQL/Conductor recovery and Inspect mutation/control/localization/top-up checks. Reopen conditions are named inputs; weekday/capacity gates are execution constraints. Historical verdict/run joins remain historical; candidate:openai-skills is RETIRED with openai/plugins named as successor. All seven layers remain SCHEDULED; no fresh E2E, model run, installation or winner adoption is claimed.

All seven preregistration hashes re-freeze together and match this table, the README, index and terminal records. Static checks qualify 12 public input digests, 251 JSON pointers and all 21 survivor/missing-vote dispositions; 18 scenarios have 50 explicitly typed arms. 82 native protocol/final-stack/evidence tests pass; prepared source bytes are unchanged by the carry. Post-carry validate.py passes (70 components,11753 files,4 profiles,239 receipts), native final-stack/evidence checks and the exact diff-range check pass, and the two-commit gitleaks range is clean. Both family delta reads and current-head CI remain required.

Round-2 adjudication

Current head e1c885ccf2da76e5f1a1ba78816a4e1738d1dd40 carries ONE fix commit on 243c2c26, without a rebase, under the CC's round-2 ruling. All seven comparisons re-freeze together; their hashes match the comparison index, README, terminal rows and this body. Retained candidate acceptance gates and tie rules replace conflicting scenario gates and ties for that candidate. The seven survivor notes label substantive versus missing-vote counts and #1026 semantics.

openai/skills is retired only as a currency source for instructions-skills. Its four installed copies remain pinned at 49f948fa as "deprecated upstream, frozen copy", in use until maintained successors qualify; nothing is uninstalled. openai/plugins is not a replacement (09-26 refutation). The three disposition records and instructions comparison name the four manifest rows and W7 questions 38–41. Review and Landing scopes, Mergify account/config/native simulation inputs, the retained PR-Agent pin, current Codex control pin, stopped-sweep names, existence-shaped Conductor/pg_durable inputs and re-freeze metadata are corrected.

Local integrity: round-2 audit passes all seven digest/rule/readiness copies and four frozen-copy rows. Protocol/final-stack/evidence modules: 82 tests, exit 0, through the T2 idle wrapper. validate.py and native evidence check pass for 11,753 files; final_stack.py --check passes with existing report-only flags. Native closeout identity scan: zero hits; directory and correction-range gitleaks pass; ranged git diff --check over 8b1ef0c..e1c885c exits 0. These are local integrity and source-record checks, not scenario execution or FINAL acceptance. New-head CI and both delta reads remain required.

Follow-ups after landing (CC, 2026-10-12T09:58:25Z)

The P3s from the final reads, recorded here because the landing did not move the head:

• GPT family (fix2 at e1c885cc): regression coverage for the documentary corrections; the missing primary captures; keep the evidence for the B-b replies; the reopen wording; inherited hostname rows; qualify the archive-dependent checks.
• Claude family (micro at e1c885cc):
  • N1: the folded P3-7 residue in two non-governing fields;
  • N2: README:58, README:203-205 and body:139 do not repeat the openai-skills role scope;
  • N3: scope attribution;
  • N4: an unnamed qualifying comparison;
  • N5: the sem authority text;
  • N6: ties between challengers.
• W7 intake: this closeout's landing starts the W7 S1 pass on the seven layers' 10-10 seeds, the delta only (ruling-final-architecture-and-inbox-20261012.md §2).

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-12T09:39:35.994915Z	e1c885c	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b9ee9e6b1f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. More of your lovely PRs please.

Reviewed commit: e1c885ccf2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".