Model invocation contract

All rollout backbones use temperature 0, planner prompt v4, verifier v3, instruction v1, a six-step planning horizon, and a 100-environment-step cap. max_tokens remains model-specific.

Model	Historical backend	max_tokens	Reasoning budget	JSON request
GPT-5.5	Azure OpenAI direct	4096	512	no explicit argument
Gemini-3.1	filesystem queue	1200	512	yes
Gemini-2.5	filesystem queue	1200	512	yes
Claude-4.8	filesystem queue	1200	0	yes
Gemma-4-31B	vLLM	1200	unset	yes
Qwen3.6-27B	vLLM	4096	unset	yes
Qwen3.6-35B-A3B	vLLM	4096	unset	yes
Qwen3.5-27B	vLLM	4096	unset	yes
InternVL3.5-38B	vLLM	2048	unset	yes

The machine-readable provenance is configs/paper_table_model_invocations.json. It intentionally omits private endpoints and credentials.

OpenAI-compatible adapter

{
  "backend": "openai_compatible",
  "model": "served-model-name",
  "base_url": "http://127.0.0.1:8100/v1",
  "api_key_env": "OPENAI_API_KEY",
  "max_tokens": 4096,
  "temperature": 0.0,
  "response_format": {"type": "json_object"},
  "extra_body": {}
}

Local OpenAI-compatible endpoints need no credential. For a remote endpoint, export the named environment variable; do not store secrets in this file.

Filesystem queue adapter

The rollout atomically moves each request directory into <queue_dir>/pending/. A worker writes its response to <queue_dir>/done/<call_id>/response.json:

{"content": "{...model JSON response...}"}

The request and response directories are removed after the response is read. They are transport state and are not part of the saved episode.

For exact cost metrics, the response may additionally contain a structured usage object or the worker's Usage(...) string. The adapter accepts OpenAI-style totals as well as Gemini's separate prompt, visible candidate, and reasoning-token counts. Hidden reasoning is never folded into the paper's visible-output column. If no usage is returned, metrics.json labels the character-count fallback as an estimate.