File size: 19,669 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 | """Export finite OracleMem instances as inspectable coverage packages.
The exact-small runner keeps candidate-by-unit coverage in memory while it
solves. These helpers serialize that hidden finite instance into the same
artifact shape expected by the coverage validation protocol, so reviewers can
inspect candidate memories and the sparse coverage matrix without rerunning a
solver or calling an external API.
"""
from __future__ import annotations
import hashlib
import json
from pathlib import Path
from typing import Any, Mapping, Sequence
from .evaluate import CandidateMemory, OracleMemInstance, generate_named_distribution
SYNTHETIC_ANNOTATOR_ID = "synthetic_generator"
PACKAGE_SCHEMA_VERSION = 1
REQUIRED_PACKAGE_FILES = (
"experiences.jsonl",
"evidence_units.jsonl",
"queries.jsonl",
"candidate_memories.jsonl",
"coverage_matrix.jsonl",
"annotation_decisions.jsonl",
"candidate_generation_manifest.json",
)
def _stable_json_hash(payload: Mapping[str, Any]) -> str:
encoded = json.dumps(payload, sort_keys=True, separators=(",", ":")).encode("utf-8")
return hashlib.sha256(encoded).hexdigest()
def _file_sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def _safe_path_token(value: str) -> str:
token = "".join(char if char.isalnum() or char in "._-" else "_" for char in value)
return token.strip("._") or "instance"
def _write_jsonl(path: Path, rows: Sequence[Mapping[str, Any]]) -> None:
with path.open("w", encoding="utf-8") as handle:
for row in rows:
handle.write(json.dumps(dict(row), sort_keys=True) + "\n")
def _write_json(path: Path, payload: Mapping[str, Any]) -> None:
path.write_text(json.dumps(dict(payload), indent=2, sort_keys=True) + "\n", encoding="utf-8")
def _coverage_label(value: float) -> str:
if value >= 1.0:
return "full"
if value >= 0.75:
return "partial_strong"
if value >= 0.5:
return "partial_weak"
return "hint_only"
def _unit_kind(unit_id: str, instance: OracleMemInstance) -> str:
if unit_id in instance.current_units:
return "current_truth"
if unit_id in instance.invalidation_units:
return "invalidation"
if unit_id in instance.stale_units:
return "stale_or_superseded"
lowered = unit_id.lower()
if lowered.startswith("current:") or ":current:" in lowered:
return "current_truth"
if lowered.startswith("invalid:") or ":invalidates:" in lowered:
return "invalidation"
if lowered.startswith("stale:"):
return "stale_or_superseded"
if lowered.startswith("context:") or ":context:" in lowered:
return "context"
if "abstain" in lowered:
return "abstention"
return "synthetic_evidence"
def _unit_state(unit_id: str, instance: OracleMemInstance) -> str:
kind = _unit_kind(unit_id, instance)
if kind == "invalidation":
return "invalidates_prior"
if kind == "stale_or_superseded":
return "superseded"
if kind == "abstention":
return "missing_evidence"
return "current"
def _source_span_id(candidate: CandidateMemory) -> str:
return f"{candidate.experience_id}:synthetic_span"
def _experience_rows(instance: OracleMemInstance) -> list[dict[str, Any]]:
by_experience: dict[str, list[CandidateMemory]] = {}
for candidate in instance.candidates:
by_experience.setdefault(candidate.experience_id, []).append(candidate)
rows: list[dict[str, Any]] = []
for experience_id, candidates in sorted(
by_experience.items(),
key=lambda item: (min(candidate.time_index for candidate in item[1]), item[0]),
):
ordered = sorted(candidates, key=lambda candidate: (candidate.time_index, candidate.candidate_id))
visible_units = sorted({unit_id for candidate in ordered for unit_id in candidate.coverage})
preferred = next(
(
candidate
for candidate in ordered
if candidate.representation_type in {"raw_span", "raw"}
),
ordered[0],
)
rows.append(
{
"experience_id": experience_id,
"session_id": f"{instance.instance_id}:synthetic",
"timestamp": min(candidate.time_index for candidate in ordered),
"text": preferred.serialized,
"source_span_ids": [_source_span_id(preferred)],
"split": "synthetic",
"visible_unit_ids": visible_units,
}
)
return rows
def _unit_source_spans(unit_id: str, candidates: Sequence[CandidateMemory]) -> list[dict[str, Any]]:
spans: list[dict[str, Any]] = []
seen: set[str] = set()
for candidate in sorted(candidates, key=lambda item: (item.time_index, item.candidate_id)):
if candidate.coverage.get(unit_id, 0.0) <= 0:
continue
span_id = _source_span_id(candidate)
if span_id in seen:
continue
spans.append(
{
"span_id": span_id,
"experience_id": candidate.experience_id,
"timestamp": candidate.time_index,
"text": candidate.serialized,
}
)
seen.add(span_id)
return spans
def _evidence_unit_rows(instance: OracleMemInstance) -> list[dict[str, Any]]:
units = sorted(
set(instance.unit_weights)
| {unit_id for candidate in instance.candidates for unit_id in candidate.coverage}
)
rows: list[dict[str, Any]] = []
for unit_id in units:
spans = _unit_source_spans(unit_id, instance.candidates)
rows.append(
{
"unit_id": unit_id,
"kind": _unit_kind(unit_id, instance),
"canonical_text": unit_id,
"source_spans": spans,
"timestamp": min((span["timestamp"] for span in spans), default=0),
"state": _unit_state(unit_id, instance),
"proposition_id": unit_id.rsplit(":", 1)[0] if ":" in unit_id else unit_id,
"annotator_ids": [SYNTHETIC_ANNOTATOR_ID],
"adjudication_status": "resolved",
"unit_weight": float(instance.unit_weights.get(unit_id, 0.0)),
}
)
return rows
def _query_rows(instance: OracleMemInstance, evidence_rows: Sequence[Mapping[str, Any]]) -> list[dict[str, Any]]:
rows: list[dict[str, Any]] = []
for index, evidence in enumerate(evidence_rows):
unit_id = str(evidence["unit_id"])
source_spans = list(evidence.get("source_spans", []))
rows.append(
{
"query_id": f"{instance.instance_id}:objective_unit:{index:04d}",
"question": f"Synthetic objective probe for {unit_id}.",
"answer": str(evidence["canonical_text"]),
"category": str(evidence["kind"]),
"required_unit_ids": [unit_id],
"answer_session_ids": sorted(
{
str(span.get("experience_id", ""))
for span in source_spans
if span.get("experience_id")
}
),
"split": "synthetic",
"unit_weight": float(evidence.get("unit_weight", 0.0)),
}
)
return rows
def _candidate_rows(
instance: OracleMemInstance,
generation_config_hash: str,
) -> list[dict[str, Any]]:
rows: list[dict[str, Any]] = []
for candidate in sorted(instance.candidates, key=lambda item: item.candidate_id):
rows.append(
{
"candidate_id": candidate.candidate_id,
"experience_id": candidate.experience_id,
"candidate_group": candidate.experience_id,
"representation_type": candidate.representation_type,
"text": candidate.serialized,
"serialized": candidate.serialized,
"cost_tokens": int(candidate.cost),
"cost": int(candidate.cost),
"generator_id": candidate.generator,
"generation_config_hash": generation_config_hash,
"time_index": int(candidate.time_index),
"confidence": float(candidate.confidence),
"estimated_value": candidate.estimated_value,
"estimated_coverage": dict(sorted(candidate.estimated_coverage.items())),
"estimator_model": candidate.estimator_model,
}
)
return rows
def _coverage_rows(
instance: OracleMemInstance,
*,
include_zero_coverage: bool = False,
) -> list[dict[str, Any]]:
rows: list[dict[str, Any]] = []
unit_ids = sorted(
set(instance.unit_weights)
| {unit_id for candidate in instance.candidates for unit_id in candidate.coverage}
)
for candidate in sorted(instance.candidates, key=lambda item: item.candidate_id):
for unit_id in unit_ids:
value = float(candidate.coverage.get(unit_id, 0.0))
if value <= 0 and not include_zero_coverage:
continue
rows.append(
{
"candidate_id": candidate.candidate_id,
"experience_id": candidate.experience_id,
"candidate_group": candidate.experience_id,
"unit_id": unit_id,
"coverage": value,
"coverage_label": _coverage_label(value) if value > 0 else "none",
"rationale": (
"Synthetic OracleMem generator assigned this positive "
"candidate-by-unit coverage cell."
),
"source_span_ids": [_source_span_id(candidate)],
"annotator_ids": [SYNTHETIC_ANNOTATOR_ID],
"adjudication_status": "resolved",
}
)
return rows
def _annotation_decision_rows(instance: OracleMemInstance, coverage_rows: Sequence[Mapping[str, Any]]) -> list[dict[str, Any]]:
return [
{
"record_id": f"{instance.instance_id}:synthetic_coverage:{index:06d}",
"record_type": "coverage_cell",
"decision": "accepted",
"primary_annotator": SYNTHETIC_ANNOTATOR_ID,
"verifier": SYNTHETIC_ANNOTATOR_ID,
"adjudicator": SYNTHETIC_ANNOTATOR_ID,
"notes": (
"Automatically generated from the finite synthetic OracleMem "
"instance; no human annotation was required."
),
"candidate_id": row["candidate_id"],
"unit_id": row["unit_id"],
}
for index, row in enumerate(coverage_rows)
]
def export_coverage_package(
instance: OracleMemInstance,
out_dir: str | Path,
*,
distribution: str | None = None,
include_zero_coverage: bool = False,
) -> dict[str, str]:
"""Write a protocol-style coverage package for one finite instance.
The package is intended for synthetic/exact-small instances where coverage
is already part of the generated finite instance. It is structural evidence
that the matrix is inspectable; it does not certify a non-synthetic human
annotation package.
"""
out_path = Path(out_dir)
out_path.mkdir(parents=True, exist_ok=True)
generation_config = {
"schema_version": PACKAGE_SCHEMA_VERSION,
"instance_id": instance.instance_id,
"seed": instance.seed,
"distribution": distribution or "unknown",
"candidate_count": len(instance.candidates),
"unit_count": len(instance.unit_weights),
"generator": "oraclemem.synthetic",
}
generation_config_hash = _stable_json_hash(generation_config)
experiences = _experience_rows(instance)
evidence_units = _evidence_unit_rows(instance)
queries = _query_rows(instance, evidence_units)
candidate_memories = _candidate_rows(instance, generation_config_hash)
coverage_matrix = _coverage_rows(
instance,
include_zero_coverage=include_zero_coverage,
)
annotation_decisions = _annotation_decision_rows(instance, coverage_matrix)
paths = {
"experiences": out_path / "experiences.jsonl",
"evidence_units": out_path / "evidence_units.jsonl",
"queries": out_path / "queries.jsonl",
"candidate_memories": out_path / "candidate_memories.jsonl",
"coverage_matrix": out_path / "coverage_matrix.jsonl",
"annotation_decisions": out_path / "annotation_decisions.jsonl",
"instance": out_path / "instance.json",
}
_write_jsonl(paths["experiences"], experiences)
_write_jsonl(paths["evidence_units"], evidence_units)
_write_jsonl(paths["queries"], queries)
_write_jsonl(paths["candidate_memories"], candidate_memories)
_write_jsonl(paths["coverage_matrix"], coverage_matrix)
_write_jsonl(paths["annotation_decisions"], annotation_decisions)
_write_json(
paths["instance"],
{
"schema_version": PACKAGE_SCHEMA_VERSION,
"instance_id": instance.instance_id,
"seed": instance.seed,
"distribution": distribution or "unknown",
"unit_weights": dict(sorted(instance.unit_weights.items())),
"current_units": list(instance.current_units),
"invalidation_units": list(instance.invalidation_units),
"stale_units": list(instance.stale_units),
"candidate_count": len(instance.candidates),
"coverage_cell_count": len(coverage_matrix),
"positive_coverage_cell_count": sum(
1 for row in coverage_matrix if float(row["coverage"]) > 0
),
},
)
file_hashes = {
path.name: _file_sha256(path)
for path in sorted(paths.values(), key=lambda item: item.name)
}
manifest_path = out_path / "candidate_generation_manifest.json"
_write_json(
manifest_path,
{
**generation_config,
"generation_config_hash": generation_config_hash,
"synthetic_instance": True,
"generators": [
{
"generator_id": "oraclemem.synthetic",
"description": "Deterministic local OracleMem synthetic generator.",
}
],
"allowed_inputs": [
"synthetic generator parameters",
"candidate-visible synthetic experience text",
],
"forbidden_inputs": [
"held-out query text",
"gold answers",
"required_unit_ids beyond the generated synthetic objective",
"oracle weights during deployable writer decisions",
"method labels",
],
"prompt_hashes": {},
"cost_counter": "candidate.cost serialized as cost_tokens",
"split_hashes": file_hashes,
"counts": {
"evidence_units": len(evidence_units),
"queries": len(queries),
"experiences": len(experiences),
"candidate_memories": len(candidate_memories),
"positive_coverage_rows": sum(
1 for row in coverage_matrix if float(row["coverage"]) > 0
),
"coverage_rows": len(coverage_matrix),
},
"notes": (
"Synthetic exact-small export. Coverage cells are generated by "
"code and are inspectable in coverage_matrix.jsonl."
),
},
)
file_hashes[manifest_path.name] = _file_sha256(manifest_path)
readme_path = out_path / "README.md"
readme_path.write_text(
"\n".join(
[
"# OracleMem Synthetic Coverage Package",
"",
f"Instance: `{instance.instance_id}`",
f"Distribution: `{distribution or 'unknown'}`",
f"Candidates: {len(instance.candidates)}",
f"Positive coverage cells: {len(coverage_matrix)}",
"",
"`coverage_matrix.jsonl` is sparse: missing candidate/unit cells are zero.",
"This package is inspectable synthetic coverage, not a human non-synthetic annotation certificate.",
"",
]
),
encoding="utf-8",
)
result = {
"manifest": str(manifest_path),
"package_dir": str(out_path),
"candidate_generation_manifest": str(manifest_path),
"README": str(readme_path),
}
result.update({key: str(path) for key, path in paths.items()})
return result
def export_coverage_packages(
*,
seeds: Sequence[int],
distributions: Sequence[str],
out_dir: str | Path,
normal_count: int = 3,
update_count: int = 2,
max_packages: int | None = None,
) -> dict[str, Any]:
"""Generate and export coverage packages for a benchmark seed grid."""
root = Path(out_dir)
root.mkdir(parents=True, exist_ok=True)
packages: list[dict[str, Any]] = []
for distribution in distributions:
distribution_dir = root / _safe_path_token(distribution)
for seed in seeds:
if max_packages is not None and len(packages) >= max_packages:
break
instance = generate_named_distribution(
distribution,
seed,
normal_count=normal_count,
update_count=update_count,
)
package_dir = distribution_dir / f"seed_{seed}"
paths = export_coverage_package(
instance,
package_dir,
distribution=distribution,
)
packages.append(
{
"distribution": distribution,
"seed": seed,
"instance_id": instance.instance_id,
"package_dir": paths["package_dir"],
"coverage_matrix": paths["coverage_matrix"],
"candidate_memories": paths["candidate_memories"],
"manifest": paths["candidate_generation_manifest"],
}
)
if max_packages is not None and len(packages) >= max_packages:
break
manifest_path = root / "coverage_export_manifest.json"
_write_json(
manifest_path,
{
"schema_version": PACKAGE_SCHEMA_VERSION,
"coverage_matrices_exportable": True,
"package_count": len(packages),
"packages": packages,
},
)
return {
"manifest": str(manifest_path),
"package_count": len(packages),
"packages": packages,
}
__all__ = [
"REQUIRED_PACKAGE_FILES",
"export_coverage_package",
"export_coverage_packages",
]
|