tests/test_decision.py

"""Standalone unit tests for research/decision.py (no pytest needed).

Run: .venv/bin/python tests/test_decision.py
"""
import sys
from pathlib import Path

sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from research.decision import (bucket_for, decide, weighted_tally,
                               accuracy_vs_coverage, bucket_abstention_curve)

TABLE = {"HIGH": {"acc": 0.5, "n": 20}, "MEDIUM": {"acc": 0.3, "n": 10},
         "LOW": {"acc": 0.1, "n": 40}, "cannot assess": {"acc": None, "n": 0}}


def test_unanimous_high_vote():
    votes = [{"verdict": "false", "conf": "HIGH"} for _ in range(3)]
    d = decide(votes, TABLE)
    assert d["verdict"] == "false"
    assert abs(d["p"] - 0.5) < 1e-9
    assert d["confidence"] == "MEDIUM"  # bucket_for(0.5)
    assert not d["abstained"]


def test_unanimous_low_abstains():
    votes = [{"verdict": "false", "conf": "LOW"} for _ in range(3)]
    d = decide(votes, TABLE, threshold=0.4)
    assert d["abstained"]
    assert d["verdict"] == "not enough information"


def test_conflict_attenuated():
    votes = [{"verdict": "false", "conf": "HIGH"},
             {"verdict": "true", "conf": "HIGH"}]
    d = decide(votes, TABLE, threshold=0.3)
    assert d["abstained"]  # p = 0.25 < 0.3
    assert abs(d["p"] - 0.25) < 1e-9


def test_majority_outweighs_minority():
    votes = [{"verdict": "overclaim", "conf": "HIGH"}] * 3 + \
            [{"verdict": "true", "conf": "HIGH"}]
    d = decide(votes, TABLE)
    assert d["verdict"] == "overclaim"
    assert abs(d["p"] - 0.375) < 1e-9  # (3*0.5)/4


def test_unknown_bucket_abstains_above_zero_threshold():
    votes = [{"verdict": "false", "conf": "weird"}]
    d = decide(votes, {}, threshold=0.1, unknown=0.0)
    assert d["abstained"]
    assert d["verdict"] == "not enough information"
    assert d["p"] == 0.0
    # threshold 0.0 = selective prediction off: emit verdict, zero confidence
    d0 = decide(votes, {}, threshold=0.0, unknown=0.0)
    assert not d0["abstained"] and d0["confidence"] == "cannot assess"


def test_weighted_tally():
    votes = [{"verdict": "a", "conf": "HIGH"}, {"verdict": "a", "conf": "LOW"},
             {"verdict": "b", "conf": "HIGH"}]
    per = weighted_tally(votes, TABLE)
    assert abs(per["a"] - 0.6) < 1e-9
    assert abs(per["b"] - 0.5) < 1e-9


def test_accuracy_vs_coverage_monotone():
    probes = [
        {"votes": [{"verdict": "false", "conf": "HIGH"}], "correct": True},
        {"votes": [{"verdict": "false", "conf": "HIGH"}], "correct": False},
        {"votes": [{"verdict": "true", "conf": "LOW"}], "correct": True},
    ]
    curve = accuracy_vs_coverage(probes, TABLE, thresholds=(0.0, 0.4))
    t0, t4 = curve[0], curve[1]
    assert t0["coverage"] == 1.0 and abs(t0["accuracy"] - 2 / 3) < 1e-9
    assert t4["coverage"] == 2 / 3 and t4["accuracy"] == 0.5  # LOW abstained


def test_bucket_abstention_curve():
    rows = [{"conf": "HIGH", "correct": True},
            {"conf": "HIGH", "correct": False},
            {"conf": "LOW", "correct": True}]
    curve = bucket_abstention_curve(rows, TABLE)
    assert curve[0]["coverage"] == 1.0 and abs(curve[0]["accuracy"] - 2 / 3) < 1e-9
    worst = curve[1]  # drop LOW (worst bucket)
    assert worst["coverage"] == 2 / 3 and worst["accuracy"] == 0.5


def test_bucket_for():
    assert bucket_for(0.9) == "HIGH"
    assert bucket_for(0.5) == "MEDIUM"
    assert bucket_for(0.2) == "LOW"
    assert bucket_for(0.0) == "cannot assess"


if __name__ == "__main__":
    fns = [v for k, v in sorted(globals().items()) if k.startswith("test_")]
    for fn in fns:
        fn()
        print(f"PASS {fn.__name__}")
    print(f"\n{len(fns)} tests passed")