File size: 18,201 Bytes
881f9f2 | 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 | #!/usr/bin/env python3
"""Render failure-taxonomy summary tables and SVG figures.
The script intentionally uses only the Python standard library so the failure
analysis lane can regenerate its evidence figures without notebook dependencies.
"""
from __future__ import annotations
import csv
import html
import textwrap
from collections import Counter, defaultdict
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
METRICS_DIR = ROOT / "results" / "metrics"
FIGURES_DIR = ROOT / "results" / "figures"
EVIDENCE_CSV = METRICS_DIR / "failure_evidence_table.csv"
CELL_ORDER = ["A", "B", "C", "D", "Y", "Z", "ZSD"]
STAGE_ORDER = [
"planning",
"tool selection",
"tool execution",
"verification",
"final answer",
]
MITIGATION_SPECS = [
{
"rank": "1",
"lane": "evidence_grounding_detection",
"mitigation_name": "missing_evidence_final_answer_guard",
"symptom": "missing-evidence final answer",
"target_pattern": "final answer or work order emitted after required evidence is missing, empty, or untrusted",
"primary_metric": "count of missing-evidence final answer rows after rerun",
"secondary_metrics": "judge_pass_rate, success_rate, latency_seconds_mean, mitigation_guard_triggered",
"stop_condition": "no reduction in target rows or low-value refusals without judge-pass improvement",
"notes": (
"Selected first because it is the largest recurring class in the current evidence table. "
"This is the truthfulness/accounting gate for the mitigation ladder: detect missing evidence "
"and block unsafe finalization; it does not retry."
),
"after_run": (
"pending guarded family reruns: "
"configs/mitigation/missing_evidence_guard_pe_self_ask.env; "
"configs/mitigation/missing_evidence_guard_verified_pe_self_ask.env"
),
"implementation_status": "implemented_pending_rerun",
},
{
"rank": "2",
"lane": "evidence_repair",
"mitigation_name": "missing_evidence_retry_replan_guard",
"symptom": "missing-evidence final answer",
"target_pattern": (
"missing evidence detected before finalization; evidence can potentially be repaired "
"by retrying the failed read-only step or replanning downstream steps"
),
"primary_metric": "supported_success_after_repair_rate",
"secondary_metrics": (
"judge_pass_rate, latency_seconds_mean, tool_call_count_mean, "
"mitigation_guard_triggered, repair_attempt_count"
),
"stop_condition": (
"no supported-success increase or excessive latency/tool-call growth relative to "
"detection-only guard"
),
"notes": (
"Next mitigation-ladder rung. Reuse the missing-evidence detector during execution; "
"retry the evidence-producing step with a bounded budget and replan the dependent suffix "
"before final answer / work-order creation. Implemented in the PE-family local runners; "
"reruns wait until detection-only guard rows exist."
),
"before_run": (
"pending detection-guard after-runs for family lanes: "
"Y+Self-Ask and Z+Self-Ask"
),
"after_run": (
"pending recovery family reruns: "
"configs/mitigation/missing_evidence_repair_pe_self_ask.env; "
"configs/mitigation/missing_evidence_repair_verified_pe_self_ask.env"
),
"before_status": "candidate_after_detection",
"after_status": "pending_rerun",
"owner_issue": "#64/#66",
"implementation_status": "implemented_pending_rerun",
},
{
"rank": "3",
"lane": "routing_contract",
"mitigation_name": "strict_tool_routing_contract",
"symptom": "tool routing or argument-contract failure",
"target_pattern": "bad tool aliases, invalid arguments, or routing-contract breaks",
"primary_metric": "count of routing or argument-contract rows after rerun",
"secondary_metrics": "tool_error_count, success_rate, judge_pass_rate",
"stop_condition": "routing errors persist or failures are hidden behind clean completion bits",
"notes": "Candidate lane; parts are already visible in the ZSD hardening history.",
},
{
"rank": "4",
"lane": "evidence_sequence",
"mitigation_name": "required_evidence_sequence_guard",
"symptom": "tool-call sequencing failure",
"target_pattern": "inference, risk estimation, or work-order creation before required evidence is acquired",
"primary_metric": "count of tool-call sequencing rows after rerun",
"secondary_metrics": "history_length, failed_steps_mean, judge_pass_rate",
"stop_condition": "agent still reasons past missing required evidence",
"notes": (
"Candidate lane; overlaps with missing-evidence repair and should not be run as a "
"separate permutation until the detection and repair rungs are measured."
),
},
{
"rank": "5",
"lane": "fault_adjudication",
"mitigation_name": "explicit_fault_risk_adjudication_step",
"symptom": "under-constrained fault/risk adjudication",
"target_pattern": "fault or risk choice remains under-justified when multiple evidence sources compete",
"primary_metric": "count of under-constrained adjudication rows after rerun",
"secondary_metrics": (
"clarity_and_justification judge dimension, judge_pass_rate, "
"wrong_fault_label_count, work_order_consistency"
),
"stop_condition": "adjudication remains vague or does not cite deciding tool evidence",
"notes": (
"Spec-ready downstream lane. Evaluate after evidence detection/repair because adjudication "
"is meaningful only when the deciding evidence exists; finalization must cite concrete tool evidence."
),
"before_run": "recovery row or detection-only row with deciding evidence present",
"after_run": "future adjudication rerun after evidence gate is active",
"before_status": "candidate_after_recovery",
"after_status": "spec_ready_deferred_until_repair",
"owner_issue": "#64/#66",
"implementation_status": "spec_ready_deferred_until_repair",
},
]
def read_rows() -> list[dict[str, str]]:
with EVIDENCE_CSV.open(newline="") as f:
return list(csv.DictReader(f))
def write_csv(path: Path, fieldnames: list[str], rows: list[dict[str, object]]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
with path.open("w", newline="") as f:
writer = csv.DictWriter(f, fieldnames=fieldnames, lineterminator="\n")
writer.writeheader()
writer.writerows(rows)
def pct(numerator: int, denominator: int) -> str:
return f"{(100.0 * numerator / denominator):.1f}" if denominator else "0.0"
def xml(text: object) -> str:
return html.escape(str(text), quote=True)
def wrap_lines(text: object, width: int) -> list[str]:
lines = textwrap.wrap(str(text), width=width, break_long_words=False)
return lines or [""]
def write_taxonomy_counts(rows: list[dict[str, str]]) -> list[dict[str, object]]:
counts = Counter(r["taxonomy_label"] for r in rows)
total = sum(counts.values())
out = [
{"taxonomy_label": label, "rows": count, "percent": pct(count, total)}
for label, count in counts.most_common()
]
write_csv(
METRICS_DIR / "failure_taxonomy_counts.csv",
["taxonomy_label", "rows", "percent"],
out,
)
return out
def write_symptom_counts(rows: list[dict[str, str]]) -> list[dict[str, object]]:
counts = Counter(r["symptom"] for r in rows)
labels_by_symptom: dict[str, Counter[str]] = defaultdict(Counter)
mitigations: dict[str, Counter[str]] = defaultdict(Counter)
for row in rows:
labels_by_symptom[row["symptom"]][row["taxonomy_label"]] += 1
mitigations[row["symptom"]][row["candidate_mitigation"]] += 1
total = sum(counts.values())
out = []
for symptom, count in counts.most_common():
out.append(
{
"symptom": symptom,
"rows": count,
"percent": pct(count, total),
"dominant_taxonomy_label": labels_by_symptom[symptom].most_common(1)[0][
0
],
"candidate_mitigation": mitigations[symptom].most_common(1)[0][0],
}
)
write_csv(
METRICS_DIR / "failure_symptom_counts.csv",
[
"symptom",
"rows",
"percent",
"dominant_taxonomy_label",
"candidate_mitigation",
],
out,
)
return out
def write_stage_cell_counts(rows: list[dict[str, str]]) -> list[dict[str, object]]:
counts = Counter((r["failure_stage"], r["cell"]) for r in rows)
out = []
for stage in STAGE_ORDER:
row: dict[str, object] = {"failure_stage": stage}
total = 0
for cell in CELL_ORDER:
value = counts[(stage, cell)]
row[cell] = value
total += value
row["total"] = total
out.append(row)
write_csv(
METRICS_DIR / "failure_stage_cell_counts.csv",
["failure_stage", *CELL_ORDER, "total"],
out,
)
return out
def write_mitigation_inventory(rows: list[dict[str, str]]) -> list[dict[str, object]]:
symptom_counts = Counter(r["symptom"] for r in rows)
out = []
for spec in MITIGATION_SPECS:
evidence_rows = symptom_counts[spec["symptom"]]
out.append(
{
"lane": spec["lane"],
"mitigation_name": spec["mitigation_name"],
"before_run": spec.get(
"before_run",
f"results/metrics/failure_evidence_table.csv:symptom={spec['symptom']}",
),
"after_run": spec.get("after_run", ""),
"before_status": spec.get(
"before_status",
f"current_count={evidence_rows}",
),
"after_status": spec.get("after_status", "pending_rerun"),
"notes": spec["notes"],
"rank": spec["rank"],
"target_pattern": spec["target_pattern"],
"evidence_rows": evidence_rows,
"primary_metric": spec["primary_metric"],
"secondary_metrics": spec["secondary_metrics"],
"stop_condition": spec["stop_condition"],
"owner_issue": spec.get("owner_issue", "#64 -> #65/#66"),
"implementation_status": spec.get("implementation_status", "candidate"),
}
)
fieldnames = [
"lane",
"mitigation_name",
"before_run",
"after_run",
"before_status",
"after_status",
"notes",
"rank",
"target_pattern",
"evidence_rows",
"primary_metric",
"secondary_metrics",
"stop_condition",
"owner_issue",
"implementation_status",
]
write_csv(METRICS_DIR / "mitigation_run_inventory.csv", fieldnames, out)
return out
def svg_bar_chart(rows: list[dict[str, object]], path: Path) -> None:
width, height = 980, 460
left, top = 330, 76
bar_h, gap = 58, 22
max_count = max(int(r["rows"]) for r in rows) or 1
palette = ["#1f6f78", "#c85a3a", "#7b8f2a"]
parts = [
f'<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 {width} {height}">',
'<rect width="100%" height="100%" fill="#fbf7ef"/>',
'<text x="40" y="42" font-family="Georgia, serif" font-size="28" fill="#1b1b1b">Failure taxonomy counts</text>',
'<text x="40" y="68" font-family="Verdana, sans-serif" font-size="13" fill="#5b5147">Rows are judge-failed trials from failure_evidence_table.csv</text>',
]
for i, row in enumerate(rows):
y = top + i * (bar_h + gap)
count = int(row["rows"])
w = int((width - left - 110) * count / max_count)
color = palette[i % len(palette)]
parts.extend(
[
f'<text x="40" y="{y + 36}" font-family="Verdana, sans-serif" font-size="16" fill="#26231f">{xml(row["taxonomy_label"])}</text>',
f'<rect x="{left}" y="{y}" width="{w}" height="{bar_h}" rx="8" fill="{color}"/>',
f'<text x="{left + w + 14}" y="{y + 36}" font-family="Verdana, sans-serif" font-size="18" font-weight="700" fill="#26231f">{count}</text>',
]
)
parts.append("</svg>")
path.write_text("\n".join(parts) + "\n")
def svg_heatmap(rows: list[dict[str, object]], path: Path) -> None:
cell_w, cell_h = 84, 48
left, top = 190, 94
width = left + cell_w * len(CELL_ORDER) + 80
height = top + cell_h * len(rows) + 84
max_value = max(int(row[cell]) for row in rows for cell in CELL_ORDER) or 1
def color(value: int) -> str:
intensity = int(245 - (155 * value / max_value))
return f"rgb(255,{intensity},{intensity - 18})" if value else "#f2eadf"
parts = [
f'<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 {width} {height}">',
'<rect width="100%" height="100%" fill="#fbf7ef"/>',
'<text x="40" y="42" font-family="Georgia, serif" font-size="28" fill="#1b1b1b">Failure stage by cell</text>',
'<text x="40" y="68" font-family="Verdana, sans-serif" font-size="13" fill="#5b5147">Darker cells indicate more judge-failed rows in that stage/cell bucket</text>',
]
for j, cell in enumerate(CELL_ORDER):
x = left + j * cell_w
parts.append(
f'<text x="{x + cell_w/2}" y="{top - 18}" text-anchor="middle" font-family="Verdana, sans-serif" font-size="15" font-weight="700" fill="#26231f">{cell}</text>'
)
for i, row in enumerate(rows):
y = top + i * cell_h
parts.append(
f'<text x="40" y="{y + 30}" font-family="Verdana, sans-serif" font-size="14" fill="#26231f">{xml(row["failure_stage"])}</text>'
)
for j, cell in enumerate(CELL_ORDER):
x = left + j * cell_w
value = int(row[cell])
parts.extend(
[
f'<rect x="{x}" y="{y}" width="{cell_w - 4}" height="{cell_h - 4}" rx="5" fill="{color(value)}" stroke="#d7c6b2"/>',
f'<text x="{x + (cell_w - 4)/2}" y="{y + 29}" text-anchor="middle" font-family="Verdana, sans-serif" font-size="15" fill="#26231f">{value}</text>',
]
)
parts.append("</svg>")
path.write_text("\n".join(parts) + "\n")
def svg_mitigation_table(rows: list[dict[str, object]], path: Path) -> None:
width, row_h = 1340, 104
height = 118 + row_h * len(rows)
cols = [54, 310, 150, 570, 230]
headings = ["#", "Mitigation", "Evidence rows", "Target pattern", "Status"]
x_positions = [34]
for w in cols[:-1]:
x_positions.append(x_positions[-1] + w)
parts = [
f'<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 {width} {height}">',
'<rect width="100%" height="100%" fill="#fbf7ef"/>',
'<text x="34" y="42" font-family="Georgia, serif" font-size="28" fill="#1b1b1b">Mitigation inventory</text>',
'<text x="34" y="68" font-family="Verdana, sans-serif" font-size="13" fill="#5b5147">Five-lane ladder: detector first, repair/adjudication only after evidence gates are measured</text>',
]
header_y = 92
parts.append(
f'<rect x="24" y="{header_y - 24}" width="{width - 48}" height="38" rx="6" fill="#263238"/>'
)
for x, heading in zip(x_positions, headings):
parts.append(
f'<text x="{x}" y="{header_y}" font-family="Verdana, sans-serif" font-size="14" font-weight="700" fill="#fff7ea">{heading}</text>'
)
for i, row in enumerate(rows):
y = 118 + i * row_h
fill = "#fffaf1" if i % 2 == 0 else "#f1e6d6"
parts.append(
f'<rect x="24" y="{y - 28}" width="{width - 48}" height="{row_h - 8}" rx="6" fill="{fill}" stroke="#dbc8af"/>'
)
values = [
(row["rank"], 4, 13),
(row["mitigation_name"], 31, 13),
(row["evidence_rows"], 10, 13),
(row["target_pattern"], 70, 12),
(
{
"implemented_pending_rerun": "implemented / pending rerun",
"candidate_next": "candidate next",
"spec_ready_pending_implementation": "spec ready / pending impl",
"spec_ready_deferred_until_repair": "spec ready / deferred",
}.get(str(row["implementation_status"]), row["implementation_status"]),
24,
12,
),
]
for x, (value, wrap_width, font_size) in zip(x_positions, values):
for line_i, line in enumerate(wrap_lines(value, wrap_width)[:4]):
parts.append(
f'<text x="{x}" y="{y - 8 + line_i * 16}" font-family="Verdana, sans-serif" font-size="{font_size}" fill="#26231f">{xml(line)}</text>'
)
parts.append("</svg>")
path.write_text("\n".join(parts) + "\n")
def main() -> None:
rows = read_rows()
FIGURES_DIR.mkdir(parents=True, exist_ok=True)
taxonomy_counts = write_taxonomy_counts(rows)
write_symptom_counts(rows)
stage_cell_counts = write_stage_cell_counts(rows)
mitigation_rows = write_mitigation_inventory(rows)
svg_bar_chart(taxonomy_counts, FIGURES_DIR / "failure_taxonomy_counts.svg")
svg_heatmap(stage_cell_counts, FIGURES_DIR / "failure_stage_cell_heatmap.svg")
svg_mitigation_table(mitigation_rows, FIGURES_DIR / "mitigation_priority_table.svg")
print(
f"Rendered {len(taxonomy_counts)} taxonomy buckets, {len(stage_cell_counts)} stage rows, {len(mitigation_rows)} mitigation lanes."
)
if __name__ == "__main__":
main()
|