Definitive defaults follow-up: the local model server is Ollama, settled by the preregistered gate

[seathatflowsinourveins]

Scope

• What this PR changes: the local model server slot of the definitive manifest is settled as Ollama by the preregistered gate (state measurement, never definitive), with both gate folders as evidence; every manifest row now carries state and measurement; the test module gains the converse of the definitive rule and six negative controls kept in the repository.
• Base commit: 8b51946ee16e542e544936e19bb793114fea948e (main).
• Lane: lane:foundation. The manifest's three pinned trading rows gain an empty state key and every row gains measurement; no trading file changes. The trading lane's own update of its definitive flags is separate (one known exception is asserted by name in the tests).
• Owned paths touched: evidence/artifacts/local-model-server-gate-20261001/, evidence/artifacts/local-model-server-gate-confirmatory-20261001/ (new), evidence/artifacts/new-wsl-definitive-defaults-20261001/ (settlements.json, controls.py new; assemble_manifest.py, render_tables.py, definitive-manifest.json changed), docs/decisions/2026-10-01-new-wsl-definitive-defaults.md, tests/test_new_wsl_definitive_defaults.py, and manifests/evidence.json in the last commit only, by registration.

SOTA sources

• Gate subject: ggml-org/llama.cpp v0.5.0 (tools/server/server-chat.cpp line 277: unsupported Responses tool type 'namespace' skipped) and ollama/ollama v0.35.0.
• Client: openai/codex rust-v0.159.3, codex-rs/model-provider/src/provider.rs lines 60 to 64 and 421 to 433 (a configured provider keeps namespace_tools: true).
• Upstream state: llama.cpp issue 23229 (closed as not planned, 2026-07-04) and pull request 23235 (open draft, last updated 2026-05-17).
• Method: the decision rule and split order of docs/decisions/2026-10-01-new-wsl-definitive-defaults.md (a split is settled by the measurement its critics name; gates first), and both preregistrations in the gate folders, hashed before their first trial.
• Negative controls: the mutation practice of this repository's tests/test_new_wsl_definitive_defaults.py at 8b51946e.

Evidence-class table

Claim	Evidence class	Command / receipt
llama-server 0 of 3 and Ollama 2 of 3 at the 300-second limit; the MCP tool call completed in 0 of 3 against 3 of 3	native_proven on one workstation (CPU, one model, one prompt)	evidence/artifacts/local-model-server-gate-20261001/receipt.json
llama-server 0 of 3 and Ollama 3 of 3 at the 1,200-second limit	native_proven, same scope	evidence/artifacts/local-model-server-gate-confirmatory-20261001/receipt.json
The folders are faithful rebuilds of the lost originals	local_integration: preregistrations match their recorded hashes, harness scripts match the hashes the worker printed, runs/ regenerated by the original collector equals the retained summaries, receipts rebuilt by the original builders	REBUILD.md in each folder
The manifest applies the settlement and refuses a wrong one	synthetic (unit tests and seven negative controls)	tests/test_new_wsl_definitive_defaults.py; controls.py
Ollama works on the new WSL, throughput, the GPU path	none: not run	the rehearsal distribution

Local commands run

$ python3 evidence/artifacts/new-wsl-definitive-defaults-20261001/assemble_manifest.py && python3 evidence/artifacts/new-wsl-definitive-defaults-20261001/render_tables.py --write docs/decisions/2026-10-01-new-wsl-definitive-defaults.md
exit 0
$ python3 -B -m unittest tests.test_new_wsl_definitive_defaults
Ran 21 tests ... OK
$ python3 -B evidence/artifacts/new-wsl-definitive-defaults-20261001/controls.py .
seven controls, each killed by the test it targets; all files restored; 21 tests OK
$ python3 scripts/component_matrix.py --write && python3 scripts/new_host_grand_list.py --write
{"flip_rule_violations": 0, "rows": 32, "status": "written"} / {"status": "written", "layers": 32, "winners": 66}
$ python3 scripts/validate.py
{"components": 69, "hashed_files": 9145, "profiles": 4, "receipts": 186, "status": "passed"}

The original worktree with both gate folders was lost when the host restarted on 2026-10-02T00:49:51Z and /tmp was emptied. The receipts in this PR carry the rebuild's time stamp (2026-10-02T02:21:30Z); the original confirmatory receipt (sha256 93789091…) is not recoverable. The manifest change was written by a GPT-6.1 Sol builder from a written brief and checked by the tests and controls above; the gate verdict was re-read from the receipts by two Claude sessions before the restart.

Decision record

docs/decisions/2026-10-01-new-wsl-definitive-defaults.md: the "Local model server" paragraph (both gates side by side, basis, scope, limits, overturn), "What follows" item 2, and "Amendment (2026-10-02)". Overturn: a llama.cpp release that passes the same gate, or a throughput comparison once both arms pass.

Host evidence

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

Checklist

• No new or changed GitHub Actions or workflows.
• No secrets are printed, logged or committed.
• No new paid hosting, subscription or billing surface.
• Peer-owned untracked files and worktrees were preserved.

Review round (2026-10-02)

An independent Claude read of 6e0e6cc4 returned 2 blocking and 7 non-blocking findings; all nine are repaired at ca363927 (main 6080214e merged in).

• Blocking 1: the sixteen per-run event logs that both receipts inventory were excluded by the *.jsonl ignore rule. They are committed and registered; all 72 and 75 inventoried files are present with the recorded hashes; a scan of the sixteen logs with the repository's private-content patterns found nothing.
• Blocking 2: the record contradicted itself on what is split and on whether anything was measured. Corrected.
• Non-blocking 3 to 9: statements without a committed source replaced by what the receipts and the compact file say; the settled row shows both families' original picks and says that neither arm passed the first gate's frozen rule; the order is attributed to the coordinator's preregistration; tests pin the settled slot set; a seventh control covers an empty settlements file; REBUILD.md rows corrected.

The repair was built by a GPT-6.1 Sol worker from the review's findings and verified by the coordinator (rebuild reproduces byte-identically; 21 tests; seven controls killed; each rewritten statement checked against the committed files).

🤖 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.