File size: 22,557 Bytes
6c5f29f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 | """Stdlib tests for the OracleMem exact-small MVP."""
from __future__ import annotations
import json
import tempfile
import unittest
from pathlib import Path
from oraclemem import evaluate
class OracleMemEvaluationTests(unittest.TestCase):
def test_exact_solver_matches_bruteforce(self) -> None:
instance = evaluate.generate_synthetic_instance(11, normal_count=2, update_count=1)
for budget in (3, 5, 8):
exact = evaluate.exact_solve(instance, budget)
brute = evaluate.brute_force_solve(instance, budget)
self.assertAlmostEqual(exact.objective_value, brute.objective_value, places=9)
self.assertTrue(
evaluate.is_feasible(instance.candidates, exact.selected_candidate_ids, budget)
)
def test_budget_and_group_feasibility_for_all_methods(self) -> None:
instance = evaluate.generate_synthetic_instance(3, normal_count=2, update_count=1)
rows = evaluate.evaluate_instance(instance, budgets=(3, 5, 7))
self.assertGreater(len(rows), 0)
for row in rows:
self.assertTrue(row.budget_feasible, row.to_json())
self.assertTrue(row.group_feasible, row.to_json())
self.assertLessEqual(row.selected_cost, row.budget)
def test_ratio_labels_are_exact_and_not_reference_denominators(self) -> None:
instance = evaluate.generate_synthetic_instance(5, normal_count=2, update_count=1)
rows = evaluate.evaluate_instance(
instance, budgets=(4,), methods=("opt", "oracle_gvt", "greedy")
)
for row in rows:
self.assertEqual(row.denominator_label, "exact_opt")
self.assertEqual(row.upper_bound_source, "exact_opt")
self.assertIsNotNone(row.ratio_to_opt)
self.assertIsNotNone(row.ratio_to_upper_bound)
self.assertLessEqual(row.ratio_to_opt or 0.0, 1.0 + 1e-9)
if row.method == "opt":
self.assertAlmostEqual(row.ratio_to_opt or 0.0, 1.0)
summary = evaluate.aggregate_results(rows)
labels = summary["label_definitions"]
self.assertIn("exact optimum is certified", labels["ratio_to_opt"])
self.assertIn("never labeled as OPT", labels["ratio_to_reference"])
def test_tombstone_benefit_on_update_stream(self) -> None:
instance = evaluate.make_update_stress_instance()
rows = evaluate.evaluate_instance(
instance,
budgets=(3,),
methods=("oracle_gvt", "no_tombstone_gvt", "no_tombstone_opt", "fact_only"),
)
by_method = {row.method: row for row in rows}
aware = by_method["oracle_gvt"]
no_tombstone = by_method["no_tombstone_gvt"]
no_tombstone_opt = by_method["no_tombstone_opt"]
fact_only = by_method["fact_only"]
self.assertGreater(aware.objective_value, no_tombstone.objective_value)
self.assertGreater(aware.objective_value, no_tombstone_opt.objective_value)
self.assertGreater(aware.objective_value, fact_only.objective_value)
self.assertEqual(aware.update_metrics["invalidation_units_covered"], 1.0)
self.assertEqual(no_tombstone.update_metrics["invalidation_units_covered"], 0.0)
self.assertEqual(no_tombstone_opt.update_metrics["invalidation_units_covered"], 0.0)
self.assertGreater(aware.update_metrics["selected_tombstone_like"], 0.0)
def test_estimated_methods_use_estimates_not_oracle_marginals(self) -> None:
instance = evaluate.OracleMemInstance(
"estimated_policy_fixture",
(
evaluate.CandidateMemory(
"oracle_best",
"exp0",
"atomic_fact",
"FACT hidden gold evidence",
1,
{"gold": 1.0},
estimated_value=0.1,
estimator_model="fixture",
),
evaluate.CandidateMemory(
"estimated_best",
"exp0",
"summary",
"SUMMARY visibly important but actually weak",
1,
{"gold": 0.0},
estimated_value=5.0,
estimator_model="fixture",
),
),
{"gold": 1.0},
current_units=("gold",),
)
rows = evaluate.evaluate_instance(
instance,
budgets=(1,),
methods=("oracle_gvt", "estimated_gvt", "estimated_utility"),
estimator_model=evaluate.DEFAULT_ESTIMATOR_MODEL,
estimator_profile="external",
)
by_method = {row.method: row for row in rows}
self.assertEqual(by_method["oracle_gvt"].selected_candidate_ids, ("oracle_best",))
self.assertEqual(by_method["estimated_gvt"].selected_candidate_ids, ("estimated_best",))
self.assertEqual(
by_method["estimated_utility"].selected_candidate_ids,
("estimated_best",),
)
self.assertLess(by_method["estimated_gvt"].objective_value, by_method["oracle_gvt"].objective_value)
self.assertEqual(
by_method["estimated_gvt"].policy_metadata["estimator_model"],
evaluate.DEFAULT_ESTIMATOR_MODEL,
)
self.assertFalse(by_method["estimated_gvt"].policy_metadata["api_called"])
def test_estimated_methods_run_on_stress_distribution(self) -> None:
rows = evaluate.run_synthetic_benchmark(
seeds=(0, 1),
budgets=(4,),
distributions=("scope_shift_v2",),
methods=("estimated_gvt", "estimated_utility"),
estimator_model=evaluate.DEFAULT_ESTIMATOR_MODEL,
)
self.assertEqual(len(rows), 4)
for row in rows:
self.assertTrue(row.budget_feasible, row.to_json())
self.assertTrue(row.group_feasible, row.to_json())
self.assertEqual(
row.policy_metadata["estimator_model"],
evaluate.DEFAULT_ESTIMATOR_MODEL,
)
self.assertEqual(row.policy_metadata["estimator_profile"], "gemini_flash_lite_v1")
def test_noisy_estimator_profile_is_local_and_supported(self) -> None:
rows = evaluate.run_synthetic_benchmark(
seeds=(0,),
budgets=(4,),
distributions=("scope_shift_v2",),
methods=("estimated_gvt", "estimated_utility"),
estimator_model=evaluate.DEFAULT_ESTIMATOR_MODEL,
estimator_profile=evaluate.NOISY_ESTIMATOR_PROFILE,
)
self.assertEqual(len(rows), 2)
for row in rows:
self.assertTrue(row.budget_feasible, row.to_json())
self.assertTrue(row.group_feasible, row.to_json())
self.assertEqual(row.policy_metadata["estimator_profile"], evaluate.NOISY_ESTIMATOR_PROFILE)
self.assertFalse(row.policy_metadata["api_called"])
def test_train_dev_estimator_uses_visible_features_not_dev_oracle(self) -> None:
train_instance = evaluate.OracleMemInstance(
"learned_train_fixture",
(
evaluate.CandidateMemory(
"train_update_good",
"train_update",
"compound_update",
"UPDATE corrected current preference with explicit invalidation",
1,
{"train_gold": 1.0},
),
evaluate.CandidateMemory(
"train_fact_weak",
"train_fact",
"atomic_fact",
"FACT weak standalone note",
1,
{"train_weak": 0.05},
),
),
{"train_gold": 4.0, "train_weak": 1.0},
)
estimator = evaluate.train_feature_utility_estimator(
(train_instance,),
train_distributions=("fixture",),
train_seeds=(0,),
ridge=0.01,
)
dev_instance = evaluate.OracleMemInstance(
"learned_dev_fixture",
(
evaluate.CandidateMemory(
"oracle_best",
"dev_exp",
"atomic_fact",
"FACT hidden gold evidence",
1,
{"gold": 1.0},
),
evaluate.CandidateMemory(
"learned_visible",
"dev_exp",
"compound_update",
"UPDATE corrected current preference with explicit invalidation",
1,
{},
),
),
{"gold": 1.0},
current_units=("gold",),
)
rows = evaluate.evaluate_instance(
dev_instance,
budgets=(1,),
methods=("oracle_gvt", "estimated_gvt", "estimated_utility"),
estimator_model=estimator.estimator_model,
estimator_profile=evaluate.LEARNED_ESTIMATOR_PROFILE,
estimator_state=estimator,
)
by_method = {row.method: row for row in rows}
self.assertEqual(by_method["oracle_gvt"].selected_candidate_ids, ("oracle_best",))
self.assertEqual(by_method["estimated_gvt"].selected_candidate_ids, ("learned_visible",))
self.assertEqual(by_method["estimated_utility"].selected_candidate_ids, ("learned_visible",))
self.assertLess(by_method["estimated_gvt"].objective_value, by_method["oracle_gvt"].objective_value)
metadata = by_method["estimated_gvt"].policy_metadata
self.assertEqual(metadata["estimator_profile"], evaluate.LEARNED_ESTIMATOR_PROFILE)
self.assertFalse(metadata["api_called"])
self.assertTrue(metadata["trained_estimator"])
self.assertTrue(metadata["oracle_coverage_used_for_training"])
self.assertFalse(metadata["oracle_coverage_used_for_dev_decision"])
def test_synthetic_train_dev_benchmark_evaluates_only_dev_seeds(self) -> None:
rows = evaluate.run_synthetic_train_dev_benchmark(
train_seeds=(0, 1),
dev_seeds=(2, 3),
budgets=(4,),
distributions=("base",),
methods=("estimated_gvt", "estimated_utility"),
normal_count=1,
update_count=1,
estimator_ridge=0.1,
)
self.assertEqual(len(rows), 4)
self.assertEqual({row.seed for row in rows}, {2, 3})
for row in rows:
self.assertTrue(row.budget_feasible, row.to_json())
self.assertTrue(row.group_feasible, row.to_json())
self.assertEqual(row.policy_metadata["estimator_profile"], evaluate.LEARNED_ESTIMATOR_PROFILE)
self.assertEqual(row.policy_metadata["train_seeds"], [0, 1])
self.assertFalse(row.policy_metadata["oracle_coverage_used_for_dev_decision"])
def test_human_package_evaluator_accepts_trained_estimator(self) -> None:
from llm_memory_validation.evaluate_human_style_examples import evaluate_human_package
train_instance = evaluate.OracleMemInstance(
"human_learned_train_fixture",
(
evaluate.CandidateMemory(
"train_visible_good",
"train_exp",
"compound_update",
"UPDATE corrected current preference with explicit invalidation",
1,
{"train_gold": 1.0},
),
evaluate.CandidateMemory(
"train_weak",
"train_weak_exp",
"summary",
"short low-signal summary",
1,
{"train_weak": 0.01},
),
),
{"train_gold": 3.0, "train_weak": 1.0},
)
estimator = evaluate.train_feature_utility_estimator(
(train_instance,),
train_distributions=("fixture",),
train_seeds=(0,),
ridge=0.01,
)
heldout_instance = evaluate.OracleMemInstance(
"human_learned_heldout_fixture",
(
evaluate.CandidateMemory(
"oracle_best",
"heldout_exp",
"atomic_fact",
"FACT hidden gold evidence",
1,
{"gold": 1.0},
),
evaluate.CandidateMemory(
"learned_visible",
"heldout_exp",
"compound_update",
"UPDATE corrected current preference with explicit invalidation",
1,
{},
),
),
{"gold": 1.0},
current_units=("gold",),
)
rows = evaluate_human_package(
heldout_instance,
budgets=(1,),
methods=("opt", "oracle_gvt", "estimated_gvt", "estimated_utility"),
estimator_model=estimator.estimator_model,
estimator_profile=evaluate.LEARNED_ESTIMATOR_PROFILE,
estimator_state=estimator,
)
by_method = {row.method: row for row in rows}
self.assertEqual(by_method["opt"].selected_candidate_ids, ("oracle_best",))
self.assertEqual(by_method["oracle_gvt"].selected_candidate_ids, ("oracle_best",))
self.assertEqual(by_method["estimated_gvt"].selected_candidate_ids, ("learned_visible",))
self.assertEqual(by_method["estimated_utility"].selected_candidate_ids, ("learned_visible",))
metadata = by_method["estimated_gvt"].policy_metadata
self.assertTrue(metadata["trained_estimator"])
self.assertTrue(metadata["oracle_coverage_used_for_training"])
self.assertFalse(metadata["oracle_coverage_used_for_dev_decision"])
def test_method_comparisons_are_nondegenerate(self) -> None:
instance = evaluate.generate_synthetic_instance(7, normal_count=3, update_count=2)
rows = evaluate.evaluate_instance(instance, budgets=(4,), methods=evaluate.DEFAULT_METHODS)
values = {row.method: round(row.objective_value, 8) for row in rows}
self.assertGreaterEqual(len(set(values.values())), 3, values)
opt_value = values["opt"]
for method, value in values.items():
self.assertLessEqual(value, opt_value + 1e-8, method)
self.assertLess(values["recency_raw"], opt_value)
self.assertIn("reservoir_raw", values)
def test_deployable_writer_baselines_are_feasible(self) -> None:
instance = evaluate.generate_synthetic_instance(9, normal_count=3, update_count=2)
rows = evaluate.evaluate_instance(
instance,
budgets=(5,),
methods=("memgpt_tiered", "mem0_extract", "amem_graph", "amac_admission"),
)
self.assertEqual(
{row.method for row in rows},
{"memgpt_tiered", "mem0_extract", "amem_graph", "amac_admission"},
)
for row in rows:
self.assertTrue(row.budget_feasible, row.to_json())
self.assertTrue(row.group_feasible, row.to_json())
self.assertLessEqual(row.selected_cost, row.budget)
self.assertEqual(
row.policy_metadata.get("policy_family"),
"deployable_writer_baseline",
)
self.assertFalse(row.policy_metadata.get("external_service_dependencies"))
self.assertFalse(row.policy_metadata.get("oracle_coverage_used_for_decision"))
self.assertIn("proxy_for", row.policy_metadata)
self.assertRegex(
row.policy_metadata.get("limitation", ""),
r"(Local proxy only|Faithful local adapter only)",
)
def test_candidate_quality_ablations_use_filtered_candidate_pools(self) -> None:
instance = evaluate.generate_synthetic_instance(10, normal_count=3, update_count=2)
rows = evaluate.evaluate_instance(
instance,
budgets=(6,),
methods=("opt", "generic_candidate_opt", "generic_candidate_gvt", "summary_candidate_opt"),
)
by_method = {row.method: row for row in rows}
by_id = evaluate.candidates_by_id(instance.candidates)
opt_value = by_method["opt"].objective_value
for method in ("generic_candidate_opt", "generic_candidate_gvt", "summary_candidate_opt"):
self.assertLessEqual(by_method[method].objective_value, opt_value + 1e-8)
self.assertEqual(
by_method[method].policy_metadata.get("policy_family"),
"candidate_quality_ablation",
)
for candidate_id in by_method["generic_candidate_opt"].selected_candidate_ids:
self.assertIn(
by_id[candidate_id].representation_type,
evaluate.GENERIC_CANDIDATE_TYPES,
)
for candidate_id in by_method["summary_candidate_opt"].selected_candidate_ids:
self.assertEqual(by_id[candidate_id].representation_type, "summary")
def test_distribution_field_and_summary_grouping(self) -> None:
rows = evaluate.run_synthetic_benchmark(
seeds=(0, 1),
budgets=(4,),
distributions=("base",),
methods=("opt", "oracle_gvt", "reservoir_raw"),
normal_count=1,
update_count=1,
)
self.assertTrue(all(row.distribution == "base" for row in rows))
summary = evaluate.aggregate_results(rows)
self.assertEqual(summary["distributions"], ["base"])
self.assertIn("by_distribution_budget_method", summary)
self.assertTrue(
all("distribution" in row for row in summary["by_distribution_budget_method"])
)
def test_outputs_include_jsonl_json_and_markdown(self) -> None:
rows = evaluate.run_synthetic_benchmark(
seeds=(0, 1),
budgets=(3, 5),
methods=("opt", "oracle_gvt", "no_tombstone_gvt", "no_tombstone_opt"),
normal_count=1,
update_count=1,
)
with tempfile.TemporaryDirectory() as tmp:
paths = evaluate.write_benchmark_outputs(rows, tmp)
raw_path = Path(paths["raw_jsonl"])
summary_json_path = Path(paths["summary_json"])
summary_md_path = Path(paths["summary_md"])
self.assertTrue(raw_path.exists())
self.assertTrue(summary_json_path.exists())
self.assertTrue(summary_md_path.exists())
raw_rows = [
json.loads(line)
for line in raw_path.read_text(encoding="utf-8").splitlines()
if line.strip()
]
self.assertEqual(len(raw_rows), len(rows))
summary = json.loads(summary_json_path.read_text(encoding="utf-8"))
self.assertIn("by_budget_method", summary)
self.assertIn("Ratio Labels", summary_md_path.read_text(encoding="utf-8"))
def test_coverage_package_export_contains_solver_inputs(self) -> None:
instance = evaluate.generate_synthetic_instance(12, normal_count=2, update_count=1)
expected_positive_rows = sum(
1
for candidate in instance.candidates
for value in candidate.coverage.values()
if value > 0
)
with tempfile.TemporaryDirectory() as tmp:
paths = evaluate.write_coverage_package(instance, tmp)
for path in paths.values():
self.assertTrue(Path(path).exists(), path)
manifest = json.loads(Path(paths["manifest"]).read_text(encoding="utf-8"))
self.assertEqual(manifest["instance_id"], instance.instance_id)
self.assertEqual(manifest["counts"]["candidate_memories"], len(instance.candidates))
self.assertEqual(manifest["counts"]["positive_coverage_rows"], expected_positive_rows)
coverage_rows = [
json.loads(line)
for line in Path(paths["coverage_matrix"]).read_text(encoding="utf-8").splitlines()
if line.strip()
]
self.assertEqual(len(coverage_rows), expected_positive_rows)
self.assertTrue(all(0.0 < row["coverage"] <= 1.0 for row in coverage_rows))
def test_exported_coverage_package_passes_structural_audit(self) -> None:
from scripts.audit_coverage_artifacts import audit_artifact
instance = evaluate.generate_synthetic_instance(13, normal_count=2, update_count=1)
with tempfile.TemporaryDirectory() as tmp:
paths = evaluate.write_coverage_package(instance, tmp)
audit = audit_artifact(
"synthetic_package",
Path(paths["package_dir"]),
"Synthetic OracleMem coverage package.",
sample_rows=1000,
)
self.assertEqual(audit.format, "coverage_package_dir")
self.assertEqual(
audit.statuses["oracle_denominator"],
"machine-checkable coverage package",
)
self.assertEqual(
audit.statuses["coverage_matrix"],
"candidate-unit coverage present",
)
self.assertTrue(all(audit.package_files.values()), audit.package_files)
coverage_rows = [
json.loads(line)
for line in Path(paths["coverage_matrix"]).read_text(encoding="utf-8").splitlines()
if line.strip()
]
self.assertTrue(all("candidate_id" in row and "unit_id" in row for row in coverage_rows))
query_rows = [
json.loads(line)
for line in Path(paths["queries"]).read_text(encoding="utf-8").splitlines()
if line.strip()
]
self.assertGreater(len(query_rows), 0)
self.assertTrue(all(row["required_unit_ids"] for row in query_rows))
if __name__ == "__main__":
unittest.main()
|