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

.gitignore

@@ -20,6 +20,10 @@ node_modules/
 # Sanitized Codex --json event streams retained as receipt evidence (client configuration output withheld).
 !evidence/artifacts/codex-use-20260926-5/*.jsonl
 !evidence/artifacts/codex-use-20260926-6/*.jsonl
+# The model-server gates' per-run Codex event streams: the frozen pass rule is scored on them and both receipts
+# inventory them by hash.
+!evidence/artifacts/local-model-server-gate-20261001/runs/*/events.jsonl
+!evidence/artifacts/local-model-server-gate-confirmatory-20261001/runs/*/events.jsonl
 # Codex app-server hooks/list oracle answers behind scripts/adoption_status.py's Codex model (throwaway homes shown
 # as $BASE, no hash value): evidence/artifacts/adoption-status-truth-20260926/README.md.
 !evidence/artifacts/adoption-status-truth-20260926/*.jsonl

docs/decisions/2026-10-01-new-wsl-definitive-defaults.md

@@ -21,8 +21,9 @@ What a default is, what makes it definitive, and what neither is:
    undetermined). A decision-round row's is the pick of two deciders in one family.
 2. A default is definitive only when both deciders of both model families named it and both critics returned
    converged. That holds for the rows the tables print in bold, and for no first-round row.
-3. The evidence class is source review plus the published evaluations it cites. A default decided on documented
-   fit says so and claims no measured lead.
+3. The blind rounds' evidence class is source review plus the published evaluations it cites. The local model
+   server's settlement adds measured local integration checks. A default decided on documented fit says so and
+   claims no measured lead.
 4. The bar is not the same in both directions for such a default. It did not have to show a measured gain, while a
    challenger replaces it only with one: an interval that excludes zero, plus the slot's gates. Measured against
    measured, the bar is the same both ways (U11 C). An option that installs nothing is the default unless a
@@ -58,7 +59,7 @@ across both catalogs. The tables below are generated from it.
 | web-research | trafilatura; Playwright CLI | first round |
 | durable-memory | Not installed until the memory head-to-head returns (the blind round's documented-fit pick is ai-memory) | measurement, at the user's request |
 | token-efficiency | ccusage; **No context-supply layer: the usage meter only** | first round; decision round, both families |
-| observation-inference | OTel Collector Contrib; Prometheus; Loki; Grafana; Phoenix; Not installed until the deciding measurement returns (the families split between llama.cpp optional inference and ollama) | first round; decision round, split between the families: a measurement decides |
+| observation-inference | OTel Collector Contrib; Prometheus; Loki; Grafana; Phoenix; Ollama | first round; settled by the preregistered gate: neither arm passed the first gate's frozen pass rule; at the 300-second wall limit llama-server passed 0 of 3 scored runs and Ollama 2 of 3 (MCP tool calls completed in 0 of 3 and 3 of 3); at the confirmatory 1,200-second wall limit llama-server passed 0 of 3 and Ollama 3 of 3; measured on one workstation; not a merit acceptance |
 | quality-evaluation | Inspect AI; Harbor (containerized agent E2E runner); Promptfoo | first round |
 | ci-supply-chain | zizmor; attest; Syft; Dependabot; codeql-sarif; actionlint (kjanat) | first round |
 | scheduling-supervision | Dagu | first round |
@@ -98,8 +99,7 @@ across both catalogs. The tables below are generated from it.
 | | | the Claude family converged on semble and the GPT family converged on SocratiCode; until the measurement returns nothing is installed for this slot and the clients' built-in tools cover it | | | |
 | memory-owner | Not installed until the memory head-to-head returns (the blind round's documented-fit pick is ai-memory) | decided by measurement at the user's request: the best-scoring eligible system of the memory head-to-head is installed, with no protected incumbent | converged | converged | no |
 | context-supply | No context-supply layer: the usage meter only | no-install default: no challenger showed a gain on the requirement's metric with an interval excluding zero | converged | converged | yes |
-| local-model-server | Not installed until the deciding measurement returns (the families split between llama.cpp optional inference and ollama) | split between the model families; the measurement the critics named decides, before the clean install | converged | converged | no |
-| | | the Claude family converged on llama.cpp optional inference and the GPT family converged on ollama; until the measurement returns nothing is installed for this slot and the clients' built-in tools cover it | | | |
+| local-model-server | Ollama | settled by the preregistered gate: neither arm passed the first gate's frozen pass rule; at the 300-second wall limit llama-server passed 0 of 3 scored runs and Ollama 2 of 3 (MCP tool calls completed in 0 of 3 and 3 of 3); at the confirmatory 1,200-second wall limit llama-server passed 0 of 3 and Ollama 3 of 3; measured on one workstation; not a merit acceptance | converged on llama.cpp optional inference in the blind round | converged on ollama in the blind round | no |
 | container-engine | Docker Engine / Moby | default on documented fit; nothing measured separates the finalists | converged | converged | yes |
 
 <!-- tables:end -->
@@ -240,17 +240,51 @@ The points a reader needs before trusting a row:
   SocratiCode appears in no public retrieval benchmark. Each critic's overturn rule keeps its own family's pick unless a
   challenger's paired interval excludes zero, so the measurement's preregistration has to state one symmetric rule and
   its tie-break before it runs, agreed by both lanes.
-- **Local model server: split between the families; nothing installs until the measurement returns.** The Claude
-  family converged on llama.cpp and the GPT family converged on Ollama, each with two deciders and a critic. The Claude
-  critic kept llama.cpp on a narrower margin after finding, in source, that llama-server's Responses endpoint drops
-  every tool whose type is not a function, which Codex sends by default, and that Ollama handles those tools; llama.cpp
-  keeps a native metrics endpoint and no cloud route. This record does not pick between the families. The route is
-  optional under the layer's requirement, so the slot's default is "not installed" until a measurement has run. The
-  two critics name different ones. The Claude critic names a first gate, pass or fail (a Codex task that must call a
-  tool from an MCP server, on each server), and then both servers on the RTX 4090 with the same 4-bit model and one
-  embedding model at one and four concurrent sessions. The GPT critic names a paired comparison of real Claude Code,
-  Codex and embedding workloads with total task completion time as the primary metric. The gate is running on the
-  workstation; its preregistration and receipt land with its result.
+- **Local model server: Ollama, settled by the preregistered gate.** The Claude family chose llama.cpp and the GPT
+  family chose Ollama, each with two deciders and a critic. The Claude critic's rule was "First gate, pass or fail:
+  Ollama replaces llama.cpp if the Codex MCP tool call fails on llama-server (its log shows 'unsupported Responses
+  tool type 'namespace' skipped') and succeeds on Ollama." The
+  [first receipt](../../evidence/artifacts/local-model-server-gate-20261001/receipt.json) and the
+  [confirmatory receipt](../../evidence/artifacts/local-model-server-gate-confirmatory-20261001/receipt.json) show:
+
+  | Result | First gate, 300-second wall limit | Confirmatory gate, 1,200-second wall limit |
+  | --- | --- | --- |
+  | llama-server scored runs passed | 0 of 3 | 0 of 3 |
+  | Ollama scored runs passed | 2 of 3 | 3 of 3 |
+  | llama-server MCP tool calls completed | 0 of 3 | 0 of 3 |
+  | Ollama MCP tool calls completed | 3 of 3 | 3 of 3 |
+
+  Neither arm passed the first gate's frozen rule. The third Ollama run completed its tool call and then hit the
+  300-second wall limit before its final answer. The confirmatory gate changed only the wall limit to 1,200 seconds
+  and returned the replacing result. The settlement follows the order the coordinator preregistered before the
+  confirmatory run: an arm failing a gate cannot win.
+  The GPT family had already chosen Ollama; the GPT critic's own workload comparison has not run. This is a measured
+  install decision, not a merit acceptance or agreement between the families in the blind round.
+  The scope is function only: whether the server's Responses layer exposes Codex's namespaced MCP tools. It claims
+  no throughput or GPU result. Both gates ran on one shared workstation, on the CPU, with one model (Qwen3-8B
+  Q4_K_M), one prompt and three scored runs per arm; they are local integration checks, not tests on the new machine.
+  The Claude critic's required comparison was on the RTX 4090 with Ollama's context at least 64k. Both gates ran on
+  CPU with 32768 tokens of context on both arms; this model cannot provide 64k. Sampling temperature was 0.6, so the
+  runs are not deterministic. llama.cpp was release v0.5.0, build b11146; later prerelease builds were not run, and
+  PR #23235 was open when the first gate checked it.
+  Ollama 0.35.0 runs the GGUF through its bundled llama-server, so the measured difference is in the Responses layer
+  in front of it. llama-server logs the converted request rather than the raw one; the evidence for two namespace
+  tools is its warnings and Ollama counting twelve tools where llama-server kept ten. The first receipt records
+  that llama.cpp documents no Codex setup at this pin. The llama-server Codex profile mirrors Ollama's launcher
+  profile, and both arms use that launcher's model catalog. `OLLAMA_NO_CLOUD=1` and loopback binding were set, but
+  outbound traffic was not captured. Load was recorded only in the confirmatory gate, so
+  tokens per second cannot be compared across the gates; wall times and token rates give context for the wall limit
+  and are not a speed comparison. The stack manifest has no Ollama component yet. Both artifact folders were rebuilt on 2026-10-02 from the retained raw
+  runs after a host restart emptied the directory that held them; `REBUILD.md` in each folder gives every file's
+  provenance, and the original confirmatory receipt (sha256 93789091…) is lost. The confirmatory receipt records
+  that other sessions' workloads ran on the same workstation and gives the load average at each run's start.
+  The Claude critic's committed reading was of Codex main (6ece7bfc21bc, 2026-10-01): it found that a configured
+  provider defaults to `namespace_tools: true` (`codex-rs/model-provider/src/provider.rs`, lines 60 to 69 and 461 to 474).
+  The same critic recorded llama.cpp issue #23229 as closed as stale, and pull request #23235 as a draft untouched
+  since 2026-05-17. `settlements.json` retains the confirmatory receipt's limitations verbatim and hashes both
+  receipts. Both receipts are registered in `manifests/evidence.json` by this pull request; the copied limit about
+  registration describes the state when the receipt was written. A llama.cpp release that passes the same gate,
+  or a throughput comparison once both arms pass, can overturn the settlement.
 
 A slot that ends split is settled by the measurement its critics name, run on the workstation or a rehearsal
 distribution before the clean install. Until then nothing is installed for it and the clients' built-in tools cover its
@@ -310,7 +344,8 @@ retrieval adverse; medium-low confidence", and if the head-to-head installs ai-m
 ## Not covered
 
 - The GPT family has not finished the foundation's 43 first-round rows (its 21-layer run is in progress) or any trading
-  slot. Each such row stays not definitive. Code search and the local model server are split.
+  slot. Each such row stays not definitive. Code search remains split; the local model server is settled by the gates
+  above.
 - The 53 first-round rows (43 foundation, 10 trading) are one blind judge and one critic each. Of the 16 layers the
   first round settled without a comparison (15 foundation layers and the base distribution), six cite an evaluation
   result for a pick; the others rest on documentation fit and CI.
@@ -327,13 +362,23 @@ retrieval adverse; medium-low confidence", and if the head-to-head installs ai-m
 ## What follows
 
 1. The GPT family's remaining returns; the manifest is re-assembled and a slot becomes definitive when the rule holds.
-2. For the two split slots, the measurement their critics name, preregistered with one symmetric decision rule and a
-   tie-break that both lanes accept, and run on the workstation. The model server's first gate (a Codex task that
-   must call one MCP tool, on each server) is running.
+2. For the one remaining split, code search, run the measurement its critics name on the workstation, preregistered
+   with one symmetric decision rule and a tie-break that both lanes accept.
 3. The install profile and the handbook take their rows from the manifest: one default per slot, never "one of".
 4. The rehearsal distribution checks the engine default on the WSL version the new host will run.
 5. Revision 5 of the U11 record and decision 6 of the program record, both in this change.
 
+## Amendment (2026-10-02)
+
+- The local model server now defaults to Ollama after the confirmatory gate; code search is the remaining split.
+  The earlier discussion of the blind round describes that round, before this measured settlement.
+- `settlements.json` records the basis, scope, receipt hashes, verbatim limits and overturn condition. The assembler
+  checks that settlements name eligible slots and that their receipt files exist with the recorded hashes.
+- Every manifest row now carries `state` and `measurement`. The returned model-server measurement is separate from
+  the pending memory and code-search measurements, and the manifest's counts keep their existing definition.
+- The tables use the settled default and label. Tests check both directions of the convergence rule, retain the
+  exact trading flag exception, and seven negative controls regenerate, check their named failure and restore all files.
+
 ## Mistakes recorded
 
 Four rows go into the anti-pattern log of `docs/harness-defaults.md` with this record: presenting open comparisons as
@@ -343,8 +388,10 @@ blind packet.
 
 ## Evidence class
 
-Source review by model judges in two families, plus the published evaluations they cite. Nothing was installed or
-measured. The pinned and project-practice rows are not judgments at all.
+Source review by model judges in two families, plus the published evaluations they cite, supports the blind rounds.
+Those rounds installed and measured no candidate. The local model server's settlement adds two measured local
+integration checks on one workstation, recorded in the receipts linked above. The pinned and project-practice rows
+are not judgments at all.
 
 - Method files, packets, the preregistration (its pins recorded at 2026-10-01T20:51:08Z, before either family's first
   unit; the file was last written at 20:52:03Z, when the Claude run's id was added)

evidence/artifacts/local-model-server-gate-20261001/Modelfile

@@ -0,0 +1,72 @@
+FROM <STATE>/models/Qwen3-8B-Q4_K_M.gguf
+TEMPLATE """
+{{- $lastUserIdx := -1 -}}
+{{- range $idx, $msg := .Messages -}}
+{{- if eq $msg.Role "user" }}{{ $lastUserIdx = $idx }}{{ end -}}
+{{- end }}
+{{- if or .System .Tools }}<|im_start|>system
+{{ if .System }}
+{{ .System }}
+{{- end }}
+{{- if .Tools }}
+
+# Tools
+
+You may call one or more functions to assist with the user query.
+
+You are provided with function signatures within <tools></tools> XML tags:
+<tools>
+{{- range .Tools }}
+{"type": "function", "function": {{ .Function }}}
+{{- end }}
+</tools>
+
+For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
+<tool_call>
+{"name": <function-name>, "arguments": <args-json-object>}
+</tool_call>
+{{- end -}}
+<|im_end|>
+{{ end }}
+{{- range $i, $_ := .Messages }}
+{{- $last := eq (len (slice $.Messages $i)) 1 -}}
+{{- if eq .Role "user" }}<|im_start|>user
+{{ .Content }}
+{{- if and $.IsThinkSet (eq $i $lastUserIdx) }}
+   {{- if $.Think -}}
+      {{- " "}}/think
+   {{- else -}}
+      {{- " "}}/no_think
+   {{- end -}}
+{{- end }}<|im_end|>
+{{ else if eq .Role "assistant" }}<|im_start|>assistant
+{{ if (and $.IsThinkSet (and .Thinking (or $last (gt $i $lastUserIdx)))) -}}
+<think>{{ .Thinking }}</think>
+{{ end -}}
+{{ if .Content }}{{ .Content }}
+{{- else if .ToolCalls }}<tool_call>
+{{ range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}
+{{ end }}</tool_call>
+{{- end }}{{ if not $last }}<|im_end|>
+{{ end }}
+{{- else if eq .Role "tool" }}<|im_start|>user
+<tool_response>
+{{ .Content }}
+</tool_response><|im_end|>
+{{ end }}
+{{- if and (ne .Role "assistant") $last }}<|im_start|>assistant
+{{ if and $.IsThinkSet (not $.Think) -}}
+<think>
+
+</think>
+
+{{ end -}}
+{{ end }}
+{{- end }}"""
+PARAMETER repeat_penalty 1
+PARAMETER stop <|im_start|>
+PARAMETER stop <|im_end|>
+PARAMETER temperature 0.6
+PARAMETER top_k 20
+PARAMETER top_p 0.95
+PARAMETER num_ctx 32768