thresholds.json

{
 "description": "Confidence thresholds of the cascade: the harness uses the model's answer when its calibrated confidence (calibration.json applied) is >= tau, and lets the agent LLM decide otherwise. Refit them for another agent LLM, other tools or another task family.",
 "rule": ">=",
 "agent_llm": "gemini-3.5-flash",
 "tool_select": 0.44503032941063925,
 "tool_params": 0.9958731532096863,
 "fitted_on": "the agent states of a shadow-mode calibration run (150 DecisionGym tasks, seeds 5000-5149, Gemini 3.5 Flash as the agent LLM), answered by this model with calibration.json applied",
 "source": "research/results/explore/e21b_shift_cascade/cal/fits.json; harness profile research/configs/harness/explore_e21b_e12_final.yaml"
}