| """Artifact-level tests for Track B calibration and derived claims.""" |
|
|
| from __future__ import annotations |
|
|
| import importlib.util |
| import unittest |
| from pathlib import Path |
|
|
|
|
| ROOT = Path(__file__).resolve().parents[1] |
| SPEC = importlib.util.spec_from_file_location( |
| "track_b_numbers", ROOT / "scripts/generate_track_b_paper_numbers.py" |
| ) |
| assert SPEC and SPEC.loader |
| MODULE = importlib.util.module_from_spec(SPEC) |
| SPEC.loader.exec_module(MODULE) |
|
|
|
|
| class TrackBAnalysisTest(unittest.TestCase): |
| def test_all_replicate_delivered_value_identities_and_inequalities(self) -> None: |
| for dataset in MODULE.DATASETS: |
| rows = MODULE.read_jsonl(ROOT / "results" / dataset / "replicates.jsonl") |
| audit = MODULE.assert_delivered_value_score(rows, dataset) |
| self.assertEqual(audit["replicate_rows_checked"], 900) |
|
|
| def test_controller_effects_are_paired_by_seed(self) -> None: |
| for dataset in MODULE.DATASETS: |
| rows = MODULE.read_jsonl(ROOT / "results" / dataset / "replicates.jsonl") |
| effects = MODULE.paired_controller_effects(rows) |
| self.assertEqual(set(effects), {"1", "2", "5", "10", "25", "50"}) |
| self.assertTrue(all(cell["n"] == 30 for cell in effects.values())) |
|
|
| def test_primary_calibration_is_unfloored_and_above_legacy_threshold(self) -> None: |
| for dataset in MODULE.DATASETS: |
| directory = ROOT / "results" / dataset |
| manifest = MODULE.read_json(directory / "manifest.json") |
| rows = MODULE.read_jsonl(directory / "calibration.jsonl") |
| audit = MODULE.calibration_audit(rows, manifest, dataset) |
| primary = audit["primary_second_score_no_reserve_frozen"] |
| self.assertGreater(primary["applied_budget_per_campaign_cents"]["min"], 500) |
| protocol = manifest["budget_calibration"] |
| pilot = set(range(protocol["pilot_seed_start"], protocol["pilot_seed_end"] + 1)) |
| evaluation = set(range(protocol["evaluation_seed_start"], protocol["evaluation_seed_end"] + 1)) |
| self.assertTrue(pilot.isdisjoint(evaluation)) |
|
|
| def test_factorial_has_all_eight_paired_cells(self) -> None: |
| for dataset in MODULE.DATASETS: |
| directory = ROOT / "results" / dataset |
| manifest = MODULE.read_json(directory / "manifest.json") |
| rows = MODULE.read_jsonl(directory / "factorial_ablation.jsonl") |
| audit = MODULE.factorial_summary(rows, manifest, dataset) |
| self.assertEqual(len(audit["cells"]), 8) |
| self.assertTrue(all(cell["underspend_pct"]["n"] == 30 for cell in audit["cells"].values())) |
| for policy in ("calibrated", "floored_500"): |
| budgets = {row["budget_cents"] for row in rows if row["budget_policy"] == policy} |
| self.assertEqual(len(budgets), 1) |
|
|
| def test_n_and_load_sensitivities_use_one_frozen_budget(self) -> None: |
| for dataset in MODULE.DATASETS: |
| directory = ROOT / "results" / dataset |
| manifest = MODULE.read_json(directory / "manifest.json") |
| rows = MODULE.read_jsonl(directory / "sensitivity.jsonl") |
| audit = MODULE.sensitivity_summary(rows, manifest, dataset) |
| self.assertEqual(len(audit["cells"]), 6) |
| self.assertEqual(len({row["budget_cents"] for row in rows}), 1) |
|
|
|
|
| if __name__ == "__main__": |
| unittest.main() |
|
|
