Stop pinning temperature, which the current model generation refuses

[xmap]

#694 (mine) pinned temperature = 0.0 in all three chat-request builders. The reasoning was that these tasks pick from a fixed verdict set under a provider-enforced schema, so sampling variance buys nothing. That reasoning holds. The parameter does not.

The measurement

Anthropic has deprecated temperature. Measured directly against the Argo gateway on 2026-08-20, one call per model:

Model	temperature: 0.0
Sonnet 4.5, Haiku 4.5, Sonnet 4.6, Opus 4.6	accepted
Opus 4.7, Sonnet 5, Opus 5	400 `temperature` is deprecated for this model

So the pin does not steady a modern model's answer, it prevents one.

Why it looked fine

Every agent shipped today declares a model on the accepting side of that line: RunDebriefer defaults to Haiku 4.5, CautionDrafter to Sonnet 4.6. Nothing in place broke. What the pin did was silently foreclose every model upgrade, and fail illegibly when one was attempted: the refusal arrives as an invalid-request error, the debrief records DebriefDeferred, and nothing in the deferral names sampling as the cause.

Found by pointing a new RunDebriefer at Sonnet 5 on the 2-BM rehearsal and watching the walk preflight refuse on call one. It would not have been found by the test suite, which stubs the provider.

What is and is not withdrawn

Kept: the port's temperature/top_p fields, both adapters sending only what a caller set, and the trace writing it to entries_decision_inferences. The provenance work stands.

Withdrawn: the judgement that these three tasks should set a value. A caller that sets nothing records nothing, which is the honest state and was the design principle #694 argued for in the first place.

A consequence worth naming

What can be recorded about how a model was asked is bounded by what the vendor still permits to be specified, and that surface is contracting as sampling knobs give way to adaptive reasoning. Recording sampling faithfully does not arrest that. The JSR paper's reproducibility section is being updated to say so.

Gates

• pytest tests/unit tests/architecture: 45199 passed, 631 skipped
• pre-commit: pyright, tach, ruff, architecture fitness all pass

The one test asserting the pinned value now asserts request_temperature is None, with the reason in a comment rather than the bare change.

🤖 Generated with Claude Code

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
apps/api/src/cora/agent/prompts
caution_drafter.py
run_debrief.py
apps/api/src/cora/operation/adapters
_llm_decide_prompt.py
Project Total

This report was generated by python-coverage-comment-action