File size: 10,414 Bytes
4be6a52 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 | """CLI contracts exercise only offline players; no GPU or neural-model loads."""
import argparse
import hashlib
import importlib.util
import json
from pathlib import Path
import pytest
ROOT = Path(__file__).resolve().parents[1]
SPEC = importlib.util.spec_from_file_location(
"stackcraft_evaluate_cli", ROOT / "scripts/evaluate_clef.py"
)
assert SPEC is not None and SPEC.loader is not None
CLI = importlib.util.module_from_spec(SPEC)
SPEC.loader.exec_module(CLI)
def test_seed_parser_and_reserved_seed_gate() -> None:
assert CLI.parse_seeds("70,80") == (70, 80)
assert CLI.parse_seeds("30000:30200") == CLI.FINAL_TEST_SEEDS
with pytest.raises(argparse.ArgumentTypeError):
CLI.parse_seeds("1,1")
args = argparse.Namespace(seeds=(30000,), final_test=False, selection_file=None)
with pytest.raises(ValueError, match="reserved test"):
CLI.validate_selection(args, {})
def selection_fixture(tmp_path):
metrics = {
"positions": 215,
"teacher_agreement": 0.5,
"mean_nll": 1.0,
"error_rate": 0.0,
"failed_predictions": [],
"complete_probability_coverage": True,
"probability_positions": 215,
"missing_probabilities": [],
"nll_is_infinite": False,
}
candidates = []
for epoch in (1, 2):
checkpoint = tmp_path / f"checkpoint-{epoch}"
checkpoint.mkdir()
metadata = {"extra": {"epoch": epoch, "dataset_manifest_sha256": "manifest"}}
(checkpoint / "training_config.json").write_text(json.dumps(metadata))
(checkpoint / "joint_head.safetensors").write_bytes(b"fake-test-bytes-not-a-model")
hashes = CLI.checkpoint_hashes(checkpoint)
evidence = tmp_path / ("validation.json" if epoch == 1 else "validation-2.json")
evidence.write_text(
json.dumps(
{
"mode": "positions",
"split": "validation",
"positions": 215,
"dataset_manifest_sha256": "manifest",
"checkpoint_sha256": hashes,
"players": {"trained": {"metrics": {**metrics, "mean_nll": float(epoch)}}},
}
)
)
candidates.append(
{
"key": f"epoch-{epoch:02d}",
"epoch": epoch,
"checkpoint_sha256": hashes,
"checkpoint_metadata": {
"path": str(checkpoint.relative_to(tmp_path) / "training_config.json")
},
"validation_evidence": {"path": evidence.name, "sha256": CLI.file_hash(evidence)},
"ineligibility_reasons": [],
}
)
args = argparse.Namespace(
seeds=CLI.FINAL_TEST_SEEDS,
final_test=True,
max_pieces=200,
players=["base", "base-fp32", "trained", "random", "heuristic"],
max_length=4096,
selection_file=tmp_path / "selection.json",
checkpoint=tmp_path / "checkpoint-1",
)
hashes = candidates[0]["checkpoint_sha256"]
selection = {
"schema_version": 1,
"selection_split": "validation",
"checkpoint_sha256": hashes,
"test_seeds": list(CLI.FINAL_TEST_SEEDS),
"test_seeds_sha256": hashlib.sha256(
CLI.canonical_json(list(CLI.FINAL_TEST_SEEDS)).encode()
).hexdigest(),
"max_pieces": 200,
"encoding_version": CLI.ENCODING_VERSION,
"max_length": 4096,
"players": args.players,
"head_dtypes": {"base": "bfloat16", "base-fp32": "float32", "trained": "float32"},
"bootstrap_samples": 10000,
"bootstrap_seed": 2026,
"max_error_rate": 0.0,
"selected_key": "epoch-01",
"validation_metrics": metrics,
"decision_rule": "lowest validation NLL",
"candidates": candidates,
"validation_evidence": candidates[0]["validation_evidence"],
}
args.selection_file.write_text(json.dumps(selection))
return args, hashes, selection
def test_final_requires_matching_frozen_checkpoint_evidence_and_cap(tmp_path) -> None:
args, hashes, selection = selection_fixture(tmp_path)
assert CLI.validate_selection(args, hashes) == selection
args.max_pieces = 100
with pytest.raises(ValueError, match="200-piece"):
CLI.validate_selection(args, hashes)
args.max_pieces = 200
with pytest.raises(ValueError, match="checkpoint_sha256"):
CLI.validate_selection(args, {"different": "checkpoint"})
(tmp_path / "validation.json").write_text("changed")
with pytest.raises(ValueError, match="evidence file/hash"):
CLI.validate_selection(args, hashes)
def test_final_requires_precision_ablation_and_eligible_selected_checkpoint(tmp_path) -> None:
args, hashes, selection = selection_fixture(tmp_path)
args.players.remove("base-fp32")
with pytest.raises(ValueError, match="all five"):
CLI.validate_selection(args, hashes)
report = json.loads((tmp_path / "validation.json").read_text())
report["players"]["trained"]["metrics"]["mean_nll"] = None
metadata = json.loads((args.checkpoint / "training_config.json").read_text())
assert "nonfinite validation mean target NLL" in CLI.validation_ineligibility(report, metadata)
def test_completed_resume_checks_player_metadata(tmp_path, monkeypatch) -> None:
monkeypatch.setattr(CLI, "_provenance", lambda: {"source_hashes": {"fake": "test"}})
output = tmp_path / "run"
args = [
"tournament",
"--players",
"random",
"--seeds",
"70",
"--max-pieces",
"2",
"--output",
str(output),
]
assert CLI.main(args) == 0
path = output / "random/player.json"
metadata = json.loads(path.read_text())
metadata["revision"] = "changed-identity"
path.write_text(json.dumps(metadata))
with pytest.raises(SystemExit):
CLI.main(args + ["--resume"])
def test_runtime_comparison_allows_declared_head_precision_but_not_encoding() -> None:
native = {"encoding_version": "one", "dtype": "torch.bfloat16", "head_dtype": "bfloat16"}
trained = {**native, "head_dtype": "float32"}
identities = {"base": {"runtime_config": native}, "trained": {"runtime_config": trained}}
CLI.validate_neural_runtimes(identities)
trained["encoding_version"] = "two"
with pytest.raises(ValueError, match="differ"):
CLI.validate_neural_runtimes(identities)
def test_offline_tournament_resumes_without_replaying_completed_player(
tmp_path, monkeypatch
) -> None:
monkeypatch.setattr(CLI, "_provenance", lambda: {"source_hashes": {"fake": "test"}})
original = CLI._load_player
calls = []
def interrupted(name, args, hashes):
calls.append(name)
if name == "heuristic":
raise RuntimeError("test interruption before second player")
return original(name, args, hashes)
output = tmp_path / "run"
args = [
"tournament",
"--players",
"random",
"heuristic",
"--seeds",
"70,80",
"--max-pieces",
"2",
"--output",
str(output),
]
monkeypatch.setattr(CLI, "_load_player", interrupted)
with pytest.raises(RuntimeError, match="test interruption"):
CLI.main(args)
first = (output / "random/seed-70.json").read_bytes()
assert calls == ["random", "heuristic"]
def resumed(name, args, hashes):
assert name == "heuristic", "completed random player must not load again"
return original(name, args, hashes)
monkeypatch.setattr(CLI, "_load_player", resumed)
assert CLI.main(args + ["--resume"]) == 0
assert first == (output / "random/seed-70.json").read_bytes()
report = json.loads((output / "report.json").read_text())
assert len(report["episodes"]) == 4
assert not report["final_test"]
changed = args.copy()
changed[changed.index("--max-pieces") + 1] = "3"
with pytest.raises(SystemExit):
CLI.main(changed + ["--resume"])
def test_reserved_cli_rejection_happens_before_player_load(tmp_path, monkeypatch) -> None:
def forbidden(*args):
raise AssertionError("must not load a player")
monkeypatch.setattr(CLI, "_load_player", forbidden)
with pytest.raises(SystemExit):
CLI.main(
[
"tournament",
"--players",
"random",
"--seeds",
"30000",
"--max-pieces",
"200",
"--output",
str(tmp_path / "run"),
]
)
assert not (tmp_path / "run").exists()
def test_positions_evaluates_every_validation_row_and_persists_predictions(
tmp_path, monkeypatch
) -> None:
from dataclasses import asdict
from stackcraft.engine import new_game
from stackcraft.players import HeuristicPlayer, observe
dataset = tmp_path / "data"
dataset.mkdir()
rows = []
for index in range(3):
observation = observe(new_game(70 + index))
rows.append(
{
"id": str(index),
"observation": asdict(observation),
"action_id": HeuristicPlayer().choose(observation).action_id,
}
)
(dataset / "manifest.json").write_text("{}")
(dataset / "train.jsonl").write_text("")
(dataset / "validation.jsonl").write_text("".join(json.dumps(row) + "\n" for row in rows))
monkeypatch.setattr(CLI, "audit_dataset", lambda *args: None)
monkeypatch.setattr(CLI, "_provenance", lambda: {"source_hashes": {"fake": "test"}})
monkeypatch.setattr(CLI, "_release", lambda: None)
output = tmp_path / "positions"
assert (
CLI.main(
[
"positions",
"--dataset",
str(dataset),
"--players",
"heuristic",
"--output",
str(output),
]
)
== 0
)
report = json.loads((output / "report.json").read_text())
assert report["positions"] == 3
assert report["players"]["heuristic"]["metrics"]["teacher_agreement"] == 1
predictions = [
json.loads(line) for line in (output / "heuristic-positions.jsonl").read_text().splitlines()
]
assert [row["id"] for row in predictions] == ["0", "1", "2"]
|