| """The split and the leakage guard. |
| |
| These are the tests that matter most in this repo. Every claim about the |
| rubrics being untainted reduces to `assert_minable` being correct and being |
| called, so both are tested directly rather than assumed. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import json |
| import sys |
| from pathlib import Path |
|
|
| import pytest |
|
|
| ROOT = Path(__file__).resolve().parent.parent |
| sys.path.insert(0, str(ROOT)) |
|
|
| import split as split_mod |
|
|
|
|
| @pytest.fixture(scope="module") |
| def sp(): |
| return split_mod.Split(split_mod.load_manifest()) |
|
|
|
|
| def test_uses_the_frozen_development_split(sp): |
| assert len(sp.source) == 5 |
| assert len(sp.evaluation) == 15 |
|
|
|
|
| def test_split_is_disjoint(sp): |
| assert not set(sp.source) & set(sp.evaluation) |
|
|
|
|
| def test_source_fraction_is_one_quarter(sp): |
| assert sp.train_fraction == pytest.approx(0.25) |
|
|
|
|
| def test_every_tier_is_represented_in_train(): |
| """A train set missing a whole difficulty class would silently bias the |
| rubrics toward whatever complexity survived the sample.""" |
| manifest = split_mod.load_manifest() |
| tiers = {e["tier"] for e in manifest["source_competitions"]} |
| assert tiers == {"low", "medium", "high"} |
|
|
|
|
| def test_eval_competition_is_refused(sp): |
| victim = sp.evaluation[0] |
| with pytest.raises(SystemExit, match="LEAKAGE"): |
| sp.assert_minable(victim) |
|
|
|
|
| def test_every_eval_competition_is_refused(sp): |
| """Not a spot check. All held-out development competitions are unreachable.""" |
| for comp in sp.evaluation: |
| with pytest.raises(SystemExit): |
| sp.assert_minable(comp) |
|
|
|
|
| def test_every_train_competition_is_allowed(sp): |
| for comp in sp.source: |
| sp.assert_minable(comp) |
|
|
|
|
| def test_unknown_competition_is_refused(sp): |
| """Allow-list, not deny-list: a typo or a renamed slug must fail closed.""" |
| with pytest.raises(SystemExit, match="LEAKAGE GUARD"): |
| sp.assert_minable("titanic") |
| with pytest.raises(SystemExit, match="LEAKAGE GUARD"): |
| sp.assert_minable("dog-breed-identificatio") |
|
|
|
|
| def test_guard_normalizes_paths_and_case(sp): |
| comp = sp.source[0] |
| sp.assert_minable(comp.upper()) |
| sp.assert_minable(f" {comp} ") |
| sp.assert_minable(f"/some/data/dir/{comp}/") |
| victim = sp.evaluation[0] |
| with pytest.raises(SystemExit): |
| sp.assert_minable(f"/data/{victim.upper()}/prepared") |
|
|
|
|
| def test_overlapping_manifest_is_rejected(): |
| """A broken manifest must fail loudly at construction, not mine quietly.""" |
| bad = {"source_competitions": [{"id": "a"}, {"id": "b"}], |
| "evaluation_competitions": [{"id": "b"}], "frozen": True} |
| with pytest.raises(SystemExit, match="SPLIT IS BROKEN"): |
| split_mod.Split(bad) |
|
|
|
|
| def test_unfrozen_manifest_is_refused_in_strict_mode(tmp_path, monkeypatch): |
| manifest = split_mod.load_manifest() |
| manifest["frozen"] = False |
| path = tmp_path / "thawed.json" |
| path.write_text(json.dumps(manifest)) |
| monkeypatch.setattr(split_mod, "MANIFEST", path) |
| monkeypatch.setattr(split_mod, "_SPLIT", None) |
| with pytest.raises(SystemExit, match="not frozen"): |
| split_mod.get(strict=True) |
|
|
|
|
| def test_default_manifest_is_the_canonical_experiment_manifest(): |
| from shared.protocol import SPLIT_MANIFEST |
| assert split_mod.MANIFEST.resolve() == SPLIT_MANIFEST.resolve() |
|
|
|
|
| |
| def test_train_competition_is_refused_by_the_measurement_guard(sp): |
| """Scoring rubrics on the competitions they were mined from would report |
| memorization as transfer -- and it would look like a great result.""" |
| for comp in sp.source: |
| with pytest.raises(SystemExit, match="CONTAMINATED MEASUREMENT"): |
| sp.assert_evaluation_only(comp) |
|
|
|
|
| def test_every_eval_competition_is_allowed_for_measurement(sp): |
| for comp in sp.evaluation: |
| sp.assert_evaluation_only(comp) |
|
|
|
|
| def test_unknown_competition_is_refused_by_both_guards(sp): |
| with pytest.raises(SystemExit): |
| sp.assert_evaluation_only("titanic") |
| with pytest.raises(SystemExit): |
| sp.assert_minable("titanic") |
|
|
|
|
| def test_the_two_guards_are_exact_complements(sp): |
| """No competition may pass both, and none of the 75 may fail both.""" |
| for comp in sp.source + sp.evaluation: |
| passed_mine = passed_eval = False |
| try: |
| sp.assert_minable(comp); passed_mine = True |
| except SystemExit: |
| pass |
| try: |
| sp.assert_evaluation_only(comp); passed_eval = True |
| except SystemExit: |
| pass |
| assert passed_mine != passed_eval, f"{comp} passed {passed_mine=} {passed_eval=}" |
|
|
