| { |
| "schema_version": 1, |
| "audit_date": "2026-10-21", |
| "source_dataset_revision": "1ca849c2128cf9ba9090a11514378c3a8277977a", |
| "scope": "published v2 evaluation", |
| "method": "Sample saved result.config from first and last row groups of first, middle and last trace files; one available result per model per selected row group. This is not an exhaustive attempt audit.", |
| "sampled_configs": 1, |
| "evaluation_settings": [ |
| { |
| "model": "openai/coder-30b", |
| "agent": "terminus-2", |
| "context_tokens": 32768, |
| "intended_output_tokens_per_call": 8192, |
| "request_output_tokens": null, |
| "thinking": false, |
| "sampling": { |
| "temperature": 0.7, |
| "top_p": 0.8, |
| "top_k": 20, |
| "repetition_penalty": 1.05 |
| }, |
| "output_limit_status": "No explicit request cap in sampled saved configs; actual fallback cap unknown" |
| } |
| ], |
| "evidence": [ |
| { |
| "trace_file": "nl2bash_v2/traces/coder.parquet", |
| "row_group": 0, |
| "attempt_id": "0408302b-c0e9-4804-94da-07c1faba4873", |
| "model": "openai/coder-30b", |
| "agent_kwargs": { |
| "temperature": 0.7, |
| "extra_body": { |
| "top_p": 0.8, |
| "top_k": 20, |
| "repetition_penalty": 1.05, |
| "chat_template_kwargs": { |
| "enable_thinking": false |
| } |
| }, |
| "max_tokens": 8192, |
| "model_info": { |
| "max_input_tokens": 32768, |
| "input_cost_per_token": 0, |
| "output_cost_per_token": 0, |
| "max_output_tokens": 8192 |
| } |
| }, |
| "resolved_timeouts": { |
| "agent": 600.0, |
| "verifier": 600.0, |
| "agent_setup": 360.0, |
| "environment_build": 600.0, |
| "trial_attempt": null, |
| "trial_cleanup_grace": 30.0, |
| "steps": [] |
| }, |
| "agent_override_timeout_sec": null |
| } |
| ], |
| "actual_fallback_output_cap": null, |
| "findings": "Saved configurations use top-level max_tokens without llm_call_kwargs.max_tokens. The affected Terminus forwarding path does not send that top-level setting to the model API. No universal effective per-response token limit is established by this sample.", |
| "fixed_code": "config/models.py:serving_agent_kwargs forwards llm_call_kwargs.max_tokens; HTTP-request regression tests cover every current registry model; this does not change historical results." |
| } |
|
|
