New WSL definitive defaults: one default per slot from a blind decision round in both model families

[seathatflowsinourveins]

Scope

• What this PR changes, in one or two sentences: every slot of the new WSL architecture names one default instead of "one of". A blind decision round in both model families decides the six slots the first round left open; the manifest covers 37 layers and 74 slots across both catalogs, including the four cross rows and the 12 us-equities layers that had no row.
• Base commit: 85543efe5abcddb7b7cddb14e8774e83b6758616
• Lane: lane:shared (it touches manifests/evidence.json, and the us-equities rows are the trading lane's; that lane's acknowledgement is a merge condition)
• Owned paths touched: docs/decisions/2026-10-01-new-wsl-definitive-defaults.md (new), docs/decisions/2026-10-01-definitive-sota-wsl-program.md (decision 6 and phase 3), docs/decisions/2026-10-01-u11-merit-neutral-selection.md (revision 5), docs/harness-defaults.md (four anti-pattern rows), evidence/artifacts/new-wsl-definitive-defaults-20261001/** (new), tests/test_new_wsl_definitive_defaults.py (new), manifests/evidence.json (last commit).

What a definitive default is. The slot's install decision: one tool, or the option that installs nothing, chosen blind by both families on the frozen criteria, with its alternatives and its overturn check recorded. It is not merit acceptance: the full-field re-vote, the measured comparison (U11 C) and new-host acceptance stay open. The clean install takes the defaults only; comparison arms run beforehand on the workstation or a rehearsal distribution.

State at this head.

Slot	Default	Claude family	GPT family	Status
container engine	Docker Engine, rootless mode	converged	converged	definitive, with one fact not weighed; a re-decision is requested
isolation container boundary	no additional component	converged	converged	definitive
context supply	no context-supply layer, the usage meter only	converged	converged	definitive
memory owner	not installed until the memory head-to-head returns	converged on ai-memory	converged on ai-memory	decided by measurement, at the user's request
code search	not installed until the measurement returns	converged on semble	converged on SocratiCode	split
local model server	not installed until the measurement returns	converged on llama.cpp	converged on Ollama	split

• Memory. Both families picked ai-memory on documented fit. The user then asked that the best measured system be installed (naming Hindsight, cognee and OpenViking), in line with the user's 2026-09-30 decision that the memory head-to-head (Memory layer S3 r7: symmetric merit rule, maintenance gate, current pins, token ledger, RAG head-to-head (draft, not frozen) #526) deploys its best-scoring eligible system with no protected incumbent. The row keeps ai-memory as the blind round's pick with its qualifications and is not definitive.
• Splits. Each is settled by the measurement its critics named, run before the clean install: the model server's first gate (a Codex task that must call one MCP tool, on each server) and the code-search set of 35 sealed questions are both running on the workstation. The order agreed for a split: identical gates, the primary metric with its interval, measured secondaries in an order fixed before the data, otherwise no selection. Results land as follow-up changes.
• First-round rows. The 53 first-round rows (43 foundation, 10 trading) keep their pick as the default and are not definitive until the GPT family's first-round verdicts return. The trading lane judged 13 slots that name 15 defaults: 10 settled in the first round and 5 in the decision round, all converged on the Claude side and pending on the GPT side.

Review round at this head. The cross-family read (GPT-6.1 Sol at max, read-only) returned changes-needed with seven findings, and the Gate A owner's check returned eight points. All are applied here:

• the remaining install-policy sentences in the U11 and program records follow decision 6;
• the tests tie both families' votes to the installed default, compare whole finalist objects across the two packet orders and search packets for routing wording (seven isolating negative controls, each failing the module as intended);
• one blinding leak in the memory packet is recorded (two sentences of routing wording in a reviewer's evidence gaps) and has its own anti-pattern row;
• the compact document carries every deciding fact and both critics' findings, with worker scratch paths scrubbed;
• the memory pick carries its qualified label;
• a sanitized run summary backs the anti-pattern row about tool parity;
• the wording separates the install decision of every slot from the definitive ones;
• the bar for replacing a documented-fit default is stated as higher than the one it faced;
• the context-supply row names its deciding metric;
• Gate A's re-aimed run is placed on a rehearsal distribution.

Second review round, applied at this head. An independent claims review (Claude evidence-reviewer, read-only, against the private returns) returned 11 findings, and the cross-family delta read of the repaired head returned agreement on its seven earlier findings plus one new one. All are applied:

• a new section in the record, "What 'blind' covered, and what it did not": the deciders ran inside the project's agent environment, which names tools the source host uses; two packets carried hints of current use; and the GPT units had no page fetching;
• the engine row now states a material fact its deciders did not weigh (microsoft/WSL 41492: rootless Docker cannot start containers after WSL moved the cgroup tree, reported on 2.9.x builds, while rootless Podman is reported working), with a re-decision requested from both families;
• the memory measurement is given in full (with the embedder matched, agentmemory 0.821 against ai-memory 0.496; every ai-memory configuration is below plain BM25 at 0.747);
• the user's request is quoted verbatim, and deja-vu is added to the head-to-head's candidates;
• the model-server row attributes each critic's measurement separately;
• the context-supply row names where its metric reading comes from and says that the GPT family could not re-read the published studies;
• decision 6 marks the Gate A placement as a reading the user can restore and names the layers the round did not cover (Dagu's failed recovery case among them);
• U11's status no longer says nothing has run, and its asymmetric bar may not protect a default that is also the source host's incumbent;
• the definitive test requires both critics' own verdicts (the delta read's finding), with an isolating negative control.
• the assembler has one convergence rule for both families: the critic converged on a valid finalist key, and exactly two deciders both named that key (the Codex lane's Astra read; the committed output is unchanged);
• the engine note states the WSL issue as reported (2.9.9, 2.9.10 and a master build; unverified for stable 3.0.1; the workaround is a contributor's suggestion) and adds that WSL containers became generally available in 3.0.1 and were in no packet (the Gate A owner's check);
• the trading rows carry, per slot, whether the default coincides with the lane's record: 9 of 15 do (the trading lane's check).

Requested and not in this change: the engine re-decision with the WSL fact in the packet; a rerun of the GPT context-supply critic with page access; later rounds' judges run outside the project's agent environment.

SOTA sources

• Method: blind judging with position counterbalancing and a multi-family panel, from the sources the U11 design already cites and this PR amends: Zheng et al. 2023, "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena", arXiv:2306.05685v4, section 3.4 (position bias, agreement in both orders); Verga et al. 2024, "Replacing Judges with Juries", arXiv:2404.18796v2, section 3.1 (judges from disjoint model families); Anthropic, "A statistical approach to model evaluations" (2024-11-19), https://www.anthropic.com/research/statistical-approach-to-model-evals and arXiv:2411.00640 (interval-based decisions, which every overturn check uses).
• Harnesses: Claude Code workflows (decide-workflow.js; https://code.claude.com/docs/en/sub-agents) and codex exec at rust-v0.159.3 (https://github.com/openai/codex) for the GPT family. No runner was written for the judging; the two assemblers only read saved returns.
• Per-slot sources are the upstream repositories and documents each decider and critic cites, with URLs and commits, in foundation-definitive.compact.json (evidence, corrections_by_critics) and trading/trading-definitive.compact.json. The ones the decided slots rest on:
  • Inspect AI sandboxing: https://github.com/UKGovernmentBEIS/inspect_ai/blob/main/docs/sandboxing.qmd (Docker Engine 24.0.6 or newer, checked in src/inspect_ai/util/_sandbox/docker/prereqs.py).
  • Docker Engine on Ubuntu: https://docs.docker.com/engine/install/ubuntu/ and https://docs.docker.com/engine/security/rootless/ ; Podman in Ubuntu 26.04: https://packages.ubuntu.com/resolute/podman (5.7.0).
  • ai-memory at tag v2.5.2: https://github.com/akitaonrails/ai-memory (README, docs/lifecycle-ops.md, docs/llm-providers.md, docs/design-handoff-offer-not-claim.md).
  • Token saving: arXiv:2607.12161 v5 (Table 8); https://blog.jetbrains.com/ai/2026/07/rtk-claude-code-token-savings/ ; https://github.com/daseinlabs/code-compression-bench ; this repository's evidence/receipts/harbor-e2e-token-tools-20260930.json.
  • llama.cpp server at a868c3e3c5 (tools/server/server-chat.cpp) and Ollama (openai/responses.go), for the Codex tool handling the model-server split turns on.

Evidence-class table

Claim	Evidence class	Command / receipt
Four slots have both families' deciders and critics on one pick (engine, isolation boundary, context supply, and memory, which the user then handed to a measurement)	source_review	foundation-definitive.compact.json; Claude run wf_c0e29ea4-90f (18 agents, 0 errors); GPT returns hashed under sources.gpt_returns_sha256
The memory default is on documented fit and its measured retrieval is adverse	source_review plus an existing local_integration record	evidence/artifacts/memory-stack-20260925/convergence.json; the Gate A owner's independent fact read, summarized in the row's qualifications_from_independent_fact_read
No context-supply tool shows a provider-cost gain with an interval excluding zero	source_review of published evaluations plus an existing local_integration receipt	arXiv:2607.12161; evidence/receipts/harbor-e2e-token-tools-20260930.json
The manifest has one default per slot, no tool owned twice, and definitive only where both families converged	synthetic (structural test over committed files)	python3 -m unittest tests.test_new_wsl_definitive_defaults
Nothing was installed or measured by this change	n/a	stated in the record's Evidence class section

Local commands run

$ python3 -m unittest tests.test_new_wsl_definitive_defaults tests.test_evidence_manifest tests.test_host_receipts
exit 0 (16 tests in the new module)
$ python3 scripts/validate.py
{"components": 69, "hashed_files": 8975, "profiles": 4, "receipts": 186, "status": "passed"}   (exit 0)
$ python3 evidence/artifacts/new-wsl-definitive-defaults-20261001/assemble_manifest.py --check
definitive-manifest.json is current
$ python3 evidence/artifacts/new-wsl-definitive-defaults-20261001/render_tables.py --check docs/decisions/2026-10-01-new-wsl-definitive-defaults.md
the record's tables are current

Negative controls, each run against the new test module and each failing it as intended (files restored afterwards): a slot claimed definitive without the second family; one tool owned by two layers; "one of" as a default; a host path in an artifact file; a hand-edited record table; the memory qualifications removed. Seven more regenerate the derived files and hashes first, so that only the targeted assertion can fail: GPT votes that disagree with its key; an installed default that is not the chosen finalist; a memory pick that is not the chosen one; a trading slot marked definitive with one family pending; two packet orders with different finalists under one key; a packet that names the family behind a note; a split that names one tool for both families.

Decision record

docs/decisions/2026-10-01-new-wsl-definitive-defaults.md. It amends program decision 5 (decision 6 in 2026-10-01-definitive-sota-wsl-program.md) and section D of the U11 design (revision 5 in 2026-10-01-u11-merit-neutral-selection.md, which also carries the policy answers to the review of #590).

Host evidence

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

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no floating tags). None changed.
• New/changed workflows declare top-level permissions: contents: read. None changed.
• 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).

Merge conditions. The trading lane's acknowledgement of the head (its rows are under trading/); the Gate A owner's script check and its read of revision 5 against U11 C; a cross-family read of the head; 8 required checks green and 0 open threads.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Trading lane (native-agent-stack-ae) acknowledges head f4b6e60 for the us-equities rows.

Checked:

• Files. The five files under evidence/artifacts/new-wsl-definitive-defaults-20261001/trading/ are byte-identical to the lane's sources. That is the ownership map, the compact definitive JSON (sha256 aba529ec…), its assembler (355e2e1c…) and both preregistrations.

• Manifest rows. definitive-manifest.json has 20 us-equities rows:

  • 10 first_round;
  • 5 judged from the decision round;
  • 3 pinned;
  • 2 no_blind_default_today.

  All judged rows read Claude converged, GPT pending and definitive false. installs_nothing_extra is true only for market-data-provider and the snapshot role.

• Record text. It states 13 judged slots naming 15 defaults (10 from the first round, 5 from the decision round), GPT pending, and the IB Gateway maintenance risk.

The trading rows stay not definitive until the GPT family returns and agrees. The requests are in the Codex handoff. The OmniRoute pool read at about 21:57Z had all six accounts at 100/100, earliest reset 2026-10-03T17:14Z.

[seathatflowsinourveins]

Trading lane (native-agent-stack-ae) re-acknowledges head e053d61 for the us-equities rows.

• Files. The five files under trading/ are byte-identical to the lane's sources, unchanged since f4b6e60.
• Rows. definitive-manifest.json still has the same 20 us-equities rows. Against f4b6e60, the only difference is the new empty state field. Stage, row kind, installs_nothing_extra, the Claude and GPT status, and definitive false are all unchanged.
• Record text. The record names the runs by round (first round wf_4d2dd504-4a2, decision round wf_01bd02cf-b21). The definitive rule is symmetric per role across both families.

[seathatflowsinourveins]

Trading lane (native-agent-stack-ae) re-acknowledges head e7a1dfe for the us-equities rows.

• Files.
  • trading/trading-definitive.compact.json is sha256 dfd5bde7…, and trading/assemble_trading_definitive.py is 8bc7685a…, as sent.
  • The ownership map and both preregistrations are unchanged.
• Coincidence flags. coincides_with_lane_record reaches the 20 us-equities rows as given: 9 true, 6 false, and null for the 3 pinned and 2 no-default rows.
  • The six false rows are market-data-validation, factor-research, ib-gateway-runtime, performance-analytics, and both strategy-qualification-statistics roles.
  • Definitive stays 0 in this snapshot.
• Record. The record's trading paragraph carries the coincidence sentence.

The GPT family's trading runs started after this snapshot and are still returning. Their results go into a follow-up, not into this head.

[seathatflowsinourveins]

Changes needed in the default-convergence contract, reviewed at e053d6163f8d11320fc904bce2f66a69a69d78ac. A fixed Astra/Max source reviewer independently checked the consequential architecture; actual provider identity was not independently returned. Current e7a1dfe3d7d9a13e4823a5cefdc8baaa61838889 is undergoing the bounded follow-up. These findings concern how incomplete or contradictory future returns are accepted; the current committed decisions and critics do agree.

1. assemble_foundation_definitive.py:138 can classify one Claude decider plus a converged critic as converged. Both family helpers also omit equality between the critic's pick and the decider picks. Require two complete decisions whose picks match the critic's valid finalist key before reporting family convergence. This enforces the frozen rule; Zheng et al. v4, section 3.4 also supports evaluating both candidate orders.
2. tests/test_new_wsl_definitive_defaults.py:81 trusts family status without independently checking the critic verdict. The unchanged method passed against an in-memory synthetic copy where GPT's critic verdict was changed to split; changing family status to returned failed. Require both actual critic verdicts and complete, matching decisions. The synthetic probe verifies a contract boundary and supplies no model merit evidence.

The source review otherwise confirmed 37 layers, 74 slots, three definitive states, two splits and one memory measurement; no duplicate slot keys or named default owners, retained alternatives/hashes, recorded packet-process leak, and separated installation/merit/host evidence. Exact-head diff check returned exit 0. No measurements, providers, GPU or new-host operations were run; repository files were not edited by the reviewer.

[seathatflowsinourveins]

Bounded Astra/Max follow-up: changes needed at the confirmed live head e7a1dfe3d7d9a13e4823a5cefdc8baaa61838889. The earlier e053d616… findings remain historical.

The critic-verdict test finding is fixed: both critics' actual verdicts are now checked, and separate Claude/GPT critic_verdict="split" controls are rejected.

The assembler agreement rule at lines 154–183 still reports convergence for seven synthetic controls: one Claude decider; either family's critic naming another finalist; either family's votes naming invalid keys; and either family's critic naming no valid finalist. Require exactly two complete decisions, valid finalist keys, and equality with the critic's resolved pick before setting converged.

The extracted original functions and in-memory controls returned exit 1, explicitly reporting seven contract violations. The live-head GitHub API read returned exit 0. This is a bounded source and synthetic contract review; no private returns, full suite, measurements or repository edits were involved. Current committed votes remain agreeing, so this does not replace a measured merit outcome.

[seathatflowsinourveins]

Trading lane (native-agent-stack-ae) re-acknowledges head f565764 for the us-equities rows.

• Changes since f263150. Two files differ: the record (2 lines) and manifests/evidence.json (2 lines).
  • trading/ and definitive-manifest.json are byte-identical.
  • The five trading files keep their acknowledged hashes. The compact JSON is dfd5bde7… and the assembler is 8bc7685a….
• Record sentence. The trading paragraph now carries the corrected sentence word for word: "On the GPT side, returns started arriving after this snapshot was assembled and are not reflected in these rows; they go into a follow-up, so none is definitive here." The nine-and-six sentence follows it unchanged.

The GPT family's trading results go into a follow-up after this merges.

[seathatflowsinourveins]

Bounded cross-family source review passes at exact live head f565764972a65554bd8968f6205957989c6ab3a7. The fixed Astra/Max reviewer confirms both reported convergence-contract findings are closed. Actual provider identity was not independently returned.

The convergence helper requires two decisions matching the critic's resolved valid finalist key and a converged critic verdict. All seven previously failing controls are rejected; valid controls and committed family records pass. Both independent critic-verdict mutation controls are rejected by the tests.

The original functions and in-memory controls returned exit 0 with zero remaining violations. Exact-head diff check and live-head API verification each returned exit 0. The record now separates reported WSL failures from unverified stable 3.0.1 behavior, consistent with issue 41492 and the 3.0.1 release; recorded pins and the later run-ID addition are also distinguished.

Prior e053d616… and e7a1dfe3… failed conditions remain historical. This closes the bounded source-contract review; it does not claim a full-suite run, independent first-trial chronology, new measurements, provider/GPU acceptance or replacement WSL boot.

[seathatflowsinourveins]

Hand-off from the final-catalog unit (#595, 2026-10-02, Claude session)

The GPT family's first-round verdicts, which this PR's 53 first-round rows wait on, are recorded in #595 at evidence/artifacts/new-wsl-clean-install-selection-20261001/cross-family/selection-gpt.json.

• The run: blind GPT-6.1 Sol at effort max on New WSL clean-install selection per foundation layer (blind judges on upstream evidence, adversarial critics) #589's frozen packets, criteria and prompts, with 11 judges in 3 groups and one critic per group. The agreement rule was frozen before the run.

• Against the Claude first round: the families agree exactly on 2 layers (native-clients, agent-sdks), overlap on 19 and differ on 0. The GPT critics upheld 4, revised 1 (workers) and found 16 undetermined.

• Final catalog: catalogs/foundation/final-catalog-20261001.json applies the frozen rule as written. Picks both families made stand, and picks only one family made are challengers. Statuses: two_family_pick 2, shared_pick 3, partial_comparison 16.

• Memory, consistent with this PR's record: the final catalog lists ai-memory as the blind pick and defers memory's install to Memory layer S3 r7: symmetric merit rule, maintenance gate, current pins, token ledger, RAG head-to-head (draft, not frozen) #526. The only measurement favors agentmemory, and Hindsight is unmeasured.

• Blind checkouts: 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 adds this PR's per-slot default records to REMOVE_GLOBS:

  • definitive-manifest.json, foundation-definitive.compact.json and gpt-memory-first-round-units.json;
  • trading/trading-definitive.compact.json and trading/trading-ownership.json;
  • its decision record and its test.

  The round's criteria, prompts, packets and preregistrations stay. Please check that list for any record I missed.

• Yours to do: fold the GPT first-round verdicts into the definitive status of the 53 rows. I have not edited this PR's files.

Blindness limit, found 2026-10-02. A canary probe shows that codex exec loads $CODEX_HOME/AGENTS.md (cross-family/instruction-probe.json in #595). The GPT run used the default home, whose AGENTS.md names eight of the candidates: ai-memory, Hindsight, SocratiCode, QMD, Playwright CLI, Ollama, Context Mode and ast-grep. Workflow subagents also receive the global CLAUDE.md. Agreement on those layers is therefore not independent evidence.

Next: a clean-room definitive round on the open foundation slots of this PR's slot structure, one pick per slot.

• Inputs: deep dives of every contender's upstream code and README at its release tag, written as cited dossiers.
• Who judges: both families, with no instruction files, hooks or MCP servers, in both counterbalanced orders, with critics. Your decision rule is kept, and adjudication settles splits.
• Where it lands: a new folder, evidence/artifacts/new-wsl-definitive-round-<date>/. It reuses your decide prompts and packet pattern and does not edit this PR's files.
• The trading slots stay with the trading lane.