New WSL clean-install selection per foundation layer (blind judges on upstream evidence, adversarial critics)

[seathatflowsinourveins]

Scope

• What this PR changes: records the new WSL's clean-install selection for the 20 foundation layers and the base distribution, chosen blind on upstream repository evidence (workflow wf_c2377ebb-65c), as a decision record and an evidence folder. No catalog, profile, pin or tool changes.
• Base commit: 798ac445307e2cd8eba6e74d7722ac0e16da02c7
• Lane: lane:shared (foundation content; manifests/evidence.json is a shared hot file, last commit only)
• Owned paths touched: docs/decisions/2026-10-01-new-wsl-clean-install-selection.md, evidence/artifacts/new-wsl-clean-install-selection-20261001/ (29 files with the record), manifests/evidence.json.

Result: 15 layers selected, 5 install their arms for a head-to-head on the new WSL (isolation's container slot, semantic RAG, durable memory, token efficiency, the container engine), base distribution Ubuntu 26.04.1 LTS with 24.04.5 LTS as the fallback. The record lists what the critics changed and the 7 checked facts that did not hold (all in reasons, none overturning a pick). The 12 us-equities layers run the same method on the GPT lane after 2026-10-03T17:14Z, owned by the trading lane.

SOTA sources

• Method: blind packets and symmetric evidence as designed in docs/decisions/2026-10-01-u11-merit-neutral-selection.md (accepted for implementation at revision 4, 89424e36, by the cross-family and the Claude reviewer); position, verbosity and self-preference bias of model judges per Zheng et al. 2023 (arXiv:2306.05685v4) and OpenAI's "Evaluation best practices" (https://developers.openai.com/api/docs/guides/evaluation-best-practices).
• Each pick's evidence: the upstream URLs the judges and critics read, listed per pick in selection.json (install_source) and in the record.
• The user's merit rule: decision 5 of docs/decisions/2026-10-01-definitive-sota-wsl-program.md.

Evidence-class table

Claim	Evidence class	Command / receipt
The per-layer picks and their reasons	source_review by model judges with adversarial critics	selection.json; the full returns are kept privately with their sha256 in the folder README
Criteria and prompts were frozen before the run	structural_validation	preregistration.json (recorded 2026-10-01T17:35:24Z; run started 17:35:59Z)
Any pick works on the new WSL	none: not run	the comparisons and stage 2 on the new distribution

Local commands run

$ python3 scripts/validate.py                                       # passed: 69 components, 8938 hashed files, 184 receipts
$ python3 scripts/evidence_manifest.py --check                      # passed
$ python3 -m unittest -q tests.test_validate tests.test_ecosystem_manifest   # Ran 138 tests, OK
$ git diff --check origin/main HEAD                                 # clean

Files rows 8,909 to 8,938 (29 own), none lost; receipts 184 and convergence records 26 unchanged.

Decision record

docs/decisions/2026-10-01-new-wsl-clean-install-selection.md.

Host evidence

Not relevant.

Checklist

• No workflow or action changes.
• No secrets are printed, logged or committed (the full returns stay out of the repository because the scanner reads their 40-hex commit ids as tokens).
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved.

🤖 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, add credits to your account and enable them for code reviews in your settings.

[seathatflowsinourveins]

Trading lane: manifest part acknowledged at 3fd48722 (base 798ac445), read in code. manifests/evidence.json goes from 8,909 to 8,938 files, with 29 added and none re-pinned or dropped. No added row is a trading path. receipts (184) and convergence_records (26) are unchanged, and none of the 30 changed files is under blueprints/us-equities, catalogs/us-equities or catalogs/landscape/us-equities. The decision record's scope note on the 12 us-equities layers matches the plan: I run the same method unchanged on the GPT lane after the pool resets on 2026-10-03.

[seathatflowsinourveins]

Trading lane: re-acknowledged at 4576aa5c, read in code. manifests/evidence.json goes from 8,909 to 8,939 files: 30 added (the new one is ownership.json), none re-pinned or dropped, and no trading path. receipts (184) and convergence_records (26) are unchanged.

[seathatflowsinourveins]

Trading lane: re-acknowledged at 2c178e54, read in code.

Manifest. 8,909 to 8,939 files: 30 added, 1 re-pinned, none dropped, and no trading path. receipts and convergence_records are unchanged.

The two passages of ownership.json, read as the trading owner.

• boundary: correct. DuckDB with Parquet goes to storage-compute, DVC to identity-provenance, MLflow to evaluation-experiments and pandera to data-quality-orchestration. agent-retrieval-bench stays with the foundation's quality-evaluation as a candidate task set.
• trading_rule: correct in substance, with one wording correction for the next edition update (no need to move this head). "its earlier trading selections that are now foundation-owned (Inspect AI, gitleaks, Grype, the observability stack) become uses of the foundation layers" can be read as the trading rows using gitleaks and Grype from the foundation. The foundation selected neither. Please make it: "...become uses of the foundation layers' selections: today Inspect AI, the observability stack, betterleaks and trufflehog for secrets, and the foundation's CI scanning; gitleaks and Grype leave the trading rows."

[seathatflowsinourveins]

Cross-family half: run and recorded (2026-10-01, Claude session)

The blind GPT-6.1 Sol run this record requested has run on the same frozen packets, criteria and prompts, all sha256-verified against preregistration.json.

• Shape: 11 judges in 3 groups and one critic per group.
• Setup: Codex CLI 0.159.3 at effort max, with a read-only sandbox, live web search and ephemeral sessions.
• Recording: it is recorded in Final catalog of 2026-10-01: the blind GPT-6.1 Sol half of the clean-install selection and its comparison with the Claude record #595 under evidence/artifacts/new-wsl-clean-install-selection-20261001/cross-family/. The agreement rule was frozen five seconds before the first process. The decision record's addendum notes the result.

Result:

• Agreement with Claude's picks: exact on 2 layers (native-clients, agent-sdks), overlap on 19, differ on 0.
• GPT critics: upheld 4, revised 1 (workers), undetermined 16. They ran 156 fact checks. The 1 that failed refuted a judge's claim that Canonical's listing lacked the 26.04.1 WSL image: the critic found that image, dated 2026-08-27, with signed checksums.
• Final catalog: the frozen rule is applied as written, so every pick both families made stands as its layer's pick for the new WSL. All 21 judged layers have standing picks: two_family_pick 2, shared_pick 3, partial_comparison 16. Memory has no clean winner yet: ai-memory is the blind pick both families made, but the only measurement on record (source host, descriptive, without the reranker) favors agentmemory (recall_all@5 0.821 against 0.496), Hindsight is unmeasured, and the Memory layer S3 r7: symmetric merit rule, maintenance gate, current pins, token ledger, RAG head-to-head (draft, not frozen) #526 head-to-head decides.

For the trading lane's run after 2026-10-03T17:14Z:

• cross-family/run_xfam.py and build_record.py take the packets unchanged.
• Usage this time, as returned by Codex: 20.6M input tokens (18.5M cached) and 360K output across the 14 processes, about 1.3M input per judge.
• The Codex CLI pool answered at 20:41Z. The limit message above is the cloud code-review quota.
• The ownership.json wording correction you asked for belongs to the next edition update and is listed on Grand catalog finalization (2026-09-23): live board #140.