File size: 10,418 Bytes
c10636a | 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 | #!/usr/bin/env python3
"""Summarize one or more evaluator-only Agent private-gold audits."""
from __future__ import annotations
import argparse
import json
import sys
from collections import Counter, defaultdict
from collections.abc import Mapping
from pathlib import Path
from typing import Any
REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from eviscene.eval.private_gold_metrics import (
agent_private_gold_category,
agent_private_gold_category_counts,
)
SCHEMA_VERSION = "ifv-agent-test-evaluation-summary-v1"
VALID_VERDICTS = frozenset({"real", "fake"})
def _read_jsonl(path: Path) -> list[dict[str, Any]]:
rows: list[dict[str, Any]] = []
with path.open(encoding="utf-8") as handle:
for line_number, line in enumerate(handle, start=1):
if not line.strip():
continue
row = json.loads(line)
if not isinstance(row, dict):
raise ValueError(f"{path}:{line_number} must be an object")
rows.append(row)
return rows
def _write_json(path: Path, payload: Mapping[str, Any]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(
json.dumps(dict(payload), ensure_ascii=False, indent=2) + "\n",
encoding="utf-8",
)
def _write_jsonl(path: Path, rows: list[Mapping[str, Any]]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(
"".join(json.dumps(dict(row), ensure_ascii=False) + "\n" for row in rows),
encoding="utf-8",
)
def _case_id(row: Mapping[str, Any]) -> str:
value = str(
row.get("unified_case_id")
or row.get("archive_source_version_id")
or row.get("case_id")
or ""
).strip()
if not value:
raise ValueError("row lacks a stable case ID")
return value
def _parse_input(value: str) -> tuple[str, Path]:
name, separator, rendered_path = value.partition("=")
if not separator or not name.strip() or not rendered_path.strip():
raise argparse.ArgumentTypeError(
"--input must be NAME=PATH, for example old100=/tmp/audit-results.jsonl"
)
return name.strip(), Path(rendered_path.strip()).expanduser().resolve()
def _metric_payload(rows: list[Mapping[str, Any]]) -> dict[str, Any]:
completed = [
row
for row in rows
if row.get("status") == "completed"
and str(row.get("gold_verdict") or "").lower() in VALID_VERDICTS
and str(row.get("candidate_verdict") or "").lower() in VALID_VERDICTS
]
matrix = {
expected: {predicted: 0 for predicted in sorted(VALID_VERDICTS)}
for expected in sorted(VALID_VERDICTS)
}
for row in completed:
expected = str(row["gold_verdict"]).lower()
predicted = str(row["candidate_verdict"]).lower()
matrix[expected][predicted] += 1
total = sum(sum(predictions.values()) for predictions in matrix.values())
correct = sum(matrix[label][label] for label in matrix)
recall = {
label: (
matrix[label][label] / sum(matrix[label].values())
if sum(matrix[label].values())
else None
)
for label in matrix
}
available_recalls = [value for value in recall.values() if value is not None]
return {
"evaluated": total,
"correct": correct,
"accuracy": correct / total if total else None,
"balanced_accuracy": (
sum(available_recalls) / len(available_recalls)
if available_recalls
else None
),
"recall_by_gold_verdict": recall,
"confusion_gold_rows_candidate_columns": matrix,
}
def _route_breakdown(rows: list[Mapping[str, Any]]) -> dict[str, dict[str, Any]]:
groups: dict[str, list[Mapping[str, Any]]] = defaultdict(list)
for row in rows:
groups[str(row.get("construction_subroute") or "unknown")].append(row)
output: dict[str, dict[str, Any]] = {}
for route, route_rows in sorted(groups.items()):
metrics = _metric_payload(route_rows)
categories = agent_private_gold_category_counts(route_rows)
output[route] = {
"selected": len(route_rows),
"completed": sum(row.get("status") == "completed" for row in route_rows),
"engineering_or_judge_errors": sum(
row.get("status") != "completed" for row in route_rows
),
**metrics,
"three_way_private_gold_categories": categories,
}
return output
def summarize_agent_test_evaluation(
*,
inputs: list[tuple[str, Path]],
test_manifest: Path,
output_dir: Path,
expected_case_count: int | None = None,
) -> dict[str, Any]:
"""Join audit records with test metadata and calculate reported metrics."""
test_manifest = test_manifest.expanduser().resolve()
output_dir = output_dir.expanduser().resolve()
manifest_rows = _read_jsonl(test_manifest)
manifest_by_case: dict[str, dict[str, Any]] = {}
for row in manifest_rows:
case_id = _case_id(row)
if case_id in manifest_by_case:
raise ValueError(f"duplicate case ID in test manifest: {case_id}")
manifest_by_case[case_id] = row
merged: list[dict[str, Any]] = []
seen: set[str] = set()
source_counts: dict[str, int] = {}
for source_name, path in inputs:
if source_name in source_counts:
raise ValueError(f"duplicate input name: {source_name}")
source_rows = _read_jsonl(path)
source_counts[source_name] = len(source_rows)
for row in source_rows:
case_id = _case_id(row)
if case_id in seen:
raise ValueError(f"case appears in multiple audit inputs: {case_id}")
seen.add(case_id)
manifest_row = manifest_by_case.get(case_id)
if manifest_row is None:
raise ValueError(f"audit case is absent from test manifest: {case_id}")
combined = dict(row)
combined["case_id"] = case_id
combined["audit_source"] = source_name
for field in (
"construction_subroute",
"production_class",
"target_route",
"capability_cell",
"factual_status",
):
combined[field] = manifest_row.get(field)
combined["private_gold_category"] = agent_private_gold_category(combined)
merged.append(combined)
if expected_case_count is not None and len(merged) != expected_case_count:
raise ValueError(
f"expected {expected_case_count} audit rows, found {len(merged)}"
)
merged.sort(key=lambda row: row["case_id"])
status_counts = Counter(str(row.get("status") or "unknown") for row in merged)
reason_quality = Counter(
str(row.get("reason_quality") or "unknown")
for row in merged
if row.get("status") == "completed"
)
fact_alignment = Counter(
str(row.get("fact_alignment") or "unknown")
for row in merged
if row.get("status") == "completed"
)
failure_modes = Counter(
str(mode)
for row in merged
if row.get("status") == "completed"
for mode in (row.get("failure_modes") or [])
if str(mode).strip()
)
wrong_rows = [
row
for row in merged
if row.get("private_gold_category") == "wrong_verdict"
]
insufficient_rows = [
row
for row in merged
if row.get("private_gold_category") == "correct_verdict_insufficient_evidence"
]
result = {
"schema_version": SCHEMA_VERSION,
"test_manifest": str(test_manifest),
"inputs": {name: str(path) for name, path in inputs},
"input_row_counts": source_counts,
"selected": len(merged),
"status_counts": dict(status_counts),
"engineering_or_judge_errors": sum(
row.get("status") != "completed" for row in merged
),
"binary_metrics": _metric_payload(merged),
"three_way_private_gold_categories": agent_private_gold_category_counts(
merged
),
"by_construction_subroute": _route_breakdown(merged),
"failure_analysis": {
"wrong_verdict_count": len(wrong_rows),
"correct_but_insufficient_evidence_count": len(insufficient_rows),
"wrong_verdict_by_subroute": dict(
Counter(
str(row.get("construction_subroute") or "unknown")
for row in wrong_rows
)
),
"wrong_verdict_fact_alignment": dict(
Counter(str(row.get("fact_alignment") or "unknown") for row in wrong_rows)
),
"insufficient_evidence_reason_quality": dict(
Counter(
str(row.get("reason_quality") or "unknown")
for row in insufficient_rows
)
),
"all_completed_reason_quality": dict(reason_quality),
"all_completed_fact_alignment": dict(fact_alignment),
"judge_failure_modes": dict(failure_modes),
},
}
_write_jsonl(output_dir / "combined-audit-results.jsonl", merged)
result["combined_audit_results"] = str(
output_dir / "combined-audit-results.jsonl"
)
_write_json(output_dir / "summary.json", result)
return result
def main() -> int:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--input", action="append", required=True, type=_parse_input)
parser.add_argument("--test-manifest", required=True)
parser.add_argument("--output-dir", required=True)
parser.add_argument("--expected-case-count", type=int)
args = parser.parse_args()
if args.expected_case_count is not None and args.expected_case_count < 1:
parser.error("--expected-case-count must be positive")
result = summarize_agent_test_evaluation(
inputs=args.input,
test_manifest=Path(args.test_manifest),
output_dir=Path(args.output_dir),
expected_case_count=args.expected_case_count,
)
print(json.dumps(result, ensure_ascii=False, indent=2))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|