File size: 3,862 Bytes
44084af | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 | """Difficulty decisions require the declared balanced pilot, never infrastructure failures."""
import importlib.util
import json
from pathlib import Path
import sys
import pytest
SCRIPTS = Path(__file__).resolve().parents[1] / "scripts"
sys.path.insert(0, str(SCRIPTS))
spec = importlib.util.spec_from_file_location("expert_calibrate", SCRIPTS / "expert_calibrate.py")
cal = importlib.util.module_from_spec(spec)
spec.loader.exec_module(cal)
def manifest():
return {f"task-{n}": {"task_id": f"task-{n}", "family": "E1", "seed": n,
"items": [f"T1:H0{n}", f"Q1:R0{n}"]} for n in range(1, 4)}
def row(tid, who="a/model", solved=True, **extra):
return {"task_id": tid, "family": "E1", "who": who, "solved": solved, "reward": float(solved),
"failing": [], "replayed": False, **extra}
def balanced():
return [row(t, m) for t in manifest() for m in ("a/model", "b/model") for _ in range(3)]
def test_one_success_is_not_a_drop_or_a_family_decision():
out = cal.decide([row("task-1")], manifest())
assert out["tasks"]["task-1"]["verdict"] == "incomplete"
assert out["families"]["E1"]["decision"] == "collect"
assert out["families"]["E1"]["too_easy"] is None
assert out["public"]["E1"] == []
def test_complete_pilot_can_trigger_hardening():
out = cal.decide(balanced(), manifest())
assert out["families"]["E1"]["decision"] == "harden"
assert all(t["verdict"] == "drop" for t in out["tasks"].values())
def test_family_totals_cannot_hide_an_unbalanced_task_cell():
rows = balanced()
rows[0]["task_id"] = "task-2"
out = cal.decide(rows, manifest())
assert out["families"]["E1"]["decision"] == "collect"
assert out["tasks"]["task-1"]["by_model"]["a/model"] == 2
def test_infrastructure_failure_is_excluded_and_requires_replacement():
rows = balanced()
rows[0].update(solved=False, invalid="provider rate limit")
out = cal.decide(rows, manifest())
assert out["valid_trials"] == 17
assert len(out["excluded"]) == 1
assert out["families"]["E1"]["decision"] == "collect"
def test_explicit_models_exclude_spread_runs():
out = cal.decide(balanced() + [row("task-1", "small/model", False)], manifest(),
models=["a/model", "b/model"])
assert out["families"]["E1"]["decision"] == "harden"
assert out["valid_trials"] == 18
def test_one_zero_is_incomplete_and_six_zeros_require_audit():
rows = balanced()
for r in rows:
if r["task_id"] == "task-1":
r.update(solved=False, failing=[("T1", True)])
out = cal.decide(rows, manifest())
assert out["tasks"]["task-1"]["verdict"] == "audit"
assert out["tasks"]["task-1"]["top_failure"] == ("T1", 6)
assert "task-1" not in out["public"]["E1"]
assert cal.decide([rows[0]], manifest())["tasks"]["task-1"]["verdict"] == "incomplete"
@pytest.mark.parametrize("reward", [{}, {"reward": None, "solved": None}, {"reward": float("nan"), "solved": 0}])
def test_missing_or_malformed_reward_is_invalid(tmp_path, reward):
trial = tmp_path / "trial"
(trial / "verifier").mkdir(parents=True)
(trial / "result.json").write_text(json.dumps({"task_name": "task-1", "verifier_result": {"rewards": reward}}))
rows = cal.trials([str(tmp_path)], manifest(), None)
assert rows[0]["invalid"] == "missing or invalid verifier reward"
def test_ci_exposes_small_sample_uncertainty():
lo, hi = cal.wilson(1, 1)
assert lo < 0.21 and hi == 1
assert cal.wilson(0, 0) is None
def test_pooled_half_success_cannot_hide_one_model_saturating():
rows = balanced()
for r in rows:
r["solved"] = r["who"] == "a/model"
out = cal.decide(rows, manifest())
assert out["families"]["E1"]["decision"] == "audit"
assert out["families"]["E1"]["in_band"] is False
assert out["public"]["E1"] == []
|