code/self_healing_finetuning/config/thresholds.py

from __future__ import annotations


DEFAULT_THRESHOLDS = {
    "hard_gates": {
        "pii_blocked_findings_max": 0,
        "mnpi_blocked_findings_max": 0,
        "license_blocked_findings_max": 0,
        "mandatory_prompt_replay_pass_rate": 1.0,
    },
    "quality_gates": {
        "aggregate_fineval_min": 0.88,
        "financebench_regression_abs_max": 0.02,
        "convfinqa_numeric_accuracy_min": 0.85,
        "phrasebank_macro_f1_min": 0.90,
        "private_prompt_replay_pass_rate_min": 0.95,
    },
    "improvement_gates": {
        "require_prod_comparison": True,
        "aggregate_score_delta_min_abs": 0.005,
        "aggregate_score_delta_min_pct": 0.50,
        "private_replay_delta_min_abs": 0.0,
        "private_replay_delta_min_pct": 0.0,
        "no_critical_regression": True,
        "max_task_regression_abs": 0.02,
        "require_rationale": True,
    },
}


ENV_PROFILES = {
    "dev": {"aggregate_fineval_min": 0.75, "allow_mock_providers": True, "require_human_approval": False},
    "stage": {"aggregate_fineval_min": 0.85, "require_human_approval": False, "require_prod_comparison": True},
    "prod": {"aggregate_fineval_min": 0.88, "mandatory_prompt_replay_pass_rate": 1.0, "require_human_approval": True},
}


def resolve_thresholds(env: str, task: str | None = None) -> dict[str, object]:
    resolved = {k: dict(v) for k, v in DEFAULT_THRESHOLDS.items()}
    profile = ENV_PROFILES.get(env, {})
    for key, value in profile.items():
        if key in resolved["quality_gates"]:
            resolved["quality_gates"][key] = value
        else:
            resolved.setdefault("environment", {})[key] = value
    if task == "quantitative_qa":
        resolved["quality_gates"]["convfinqa_numeric_accuracy_min"] = 0.88
    if task == "sentiment":
        resolved["quality_gates"]["phrasebank_macro_f1_min"] = 0.90
    return resolved