smartgridbench-review-artifact / code /scripts /audit_failure_taxonomy.py
garn-garn's picture
Refresh SmartGridBench reviewer artifact
d606d10 verified
"""Hand-audit support for the post-PR175 failure taxonomy (#194).
PR #193 produced `results/metrics/failure_taxonomy_current.csv` with
46 rows flagged `audit_status='stratified_sample'` (one per non-empty
`(experiment_cell, auto_taxonomy_label)` stratum across paper-eligible
failures). This script is the audit-side companion: it joins each
sample row to its judge log, extracts the agent trajectory + judge
verdict, and emits a per-row brief the auditor reads to decide:
audit_decision confirmed | relabel_suggested | evidence_thin
audit_note one-sentence concrete observation
berkeley_label specification | inter_agent_orchestration | task_verification
failure_stage planning | tool_selection | tool_execution | verification | final_answer
audit_decision_source "<reviewer>:<YYYY-MM-DD>"
Subcommands:
briefs — emit per-row briefs to stdout or --out for the auditor.
By default reads both `stratified_sample` and
`manual_confirmed` rows; filter via `--status`.
add-columns — merge a hand-edited audit-decisions sidecar TSV into
the CSV (defaults to `data/audit/issue194_decisions.tsv`,
which IS committed so the audit is reproducible from
inputs). Sets `audit_status` to `manual_confirmed` for
audited rows; idempotent on re-runs. Validates
audit_decision / berkeley_label / failure_stage
against the canonical vocab AND
audit_decision_source against `<handle>:<YYYY-MM-DD>`.
Fails loudly on unmatched keys or zero updates.
render — write the Markdown audit summary by reading the
augmented CSV plus
`data/audit/issue194_recurring_patterns.json` (also
committed) for the methodology paragraph, the
Berkeley-label tie-break rule, the failure-stage
rule (separate from the Berkeley rule per
`docs/failure_taxonomy_evidence.md:139-145` vs
`:149-151`), the recurring-patterns section, and
the v2->v3->v4 tie-break-relabel callout.
Reproducible: re-running render over committed
inputs is byte-identical to the committed doc.
Mirrors `scripts/audit_failure_evidence.py` (PR #189) but targets the
new failure_taxonomy_current.csv schema instead of the preliminary
PR #151 35-row table.
"""
from __future__ import annotations
import argparse
import csv
import json
import os
import re
import sys
from collections import Counter
from pathlib import Path
REPO_ROOT = Path(__file__).resolve().parents[1]
TAXONOMY_CSV = REPO_ROOT / "results" / "metrics" / "failure_taxonomy_current.csv"
JUDGE_LOGS_ROOT = REPO_ROOT / "results" / "judge_logs"
DECISIONS_TSV = REPO_ROOT / "data" / "audit" / "issue194_decisions.tsv"
PATTERNS_JSON = REPO_ROOT / "data" / "audit" / "issue194_recurring_patterns.json"
def _long(p: Path) -> str:
"""Return a Windows long-path-prefixed string for file ops.
Some `run_name` values push the absolute judge-log path past Windows'
260-character MAX_PATH limit, which makes `Path.exists()` and
`Path.read_text()` silently report not-found. The `\\\\?\\` prefix
bypasses MAX_PATH on Win32 file APIs. POSIX is unaffected.
"""
if os.name != "nt":
return str(p)
s = str(p.resolve())
return s if s.startswith("\\\\?\\") else "\\\\?\\" + s
def _exists(p: Path) -> bool:
return os.path.exists(_long(p))
def _read_text(p: Path, encoding: str = "utf-8") -> str:
with open(_long(p), encoding=encoding) as f:
return f.read()
AUDIT_COLUMNS = [
"audit_decision",
"berkeley_label",
"failure_stage",
"audit_note",
"audit_decision_source",
]
BERKELEY_LABELS = {
"specification",
"inter_agent_orchestration",
"task_verification",
}
FAILURE_STAGES = {
"planning",
"tool_selection",
"tool_execution",
"verification",
"final_answer",
}
AUDIT_DECISIONS = {"confirmed", "relabel_suggested", "evidence_thin"}
# `audit_decision_source` should be `<reviewer-handle>:<YYYY-MM-DD>` per the
# audit-tooling docstring. Without validation, typos like `reviewer:2026-5-7`
# silently merge and break later programmatic provenance walks.
AUDIT_SOURCE_PATTERN = re.compile(r"^[a-z][a-z0-9_]*:\d{4}-\d{2}-\d{2}$")
def _judge_log_path(run_name: str, scenario_id: str, trial_index: str) -> Path:
"""Locate the judge log file for one (run, scenario, trial)."""
pad = f"{int(trial_index):02d}" if str(trial_index).isdigit() else trial_index
return JUDGE_LOGS_ROOT / run_name / f"{scenario_id}_run{pad}_judge_log.json"
def _load_taxonomy_rows() -> tuple[list[str], list[dict]]:
with TAXONOMY_CSV.open(encoding="utf-8", newline="") as f:
reader = csv.DictReader(f)
rows = list(reader)
return list(rows[0].keys()), rows
def _extract_tool_call_summary(prompt_user: str) -> list[str]:
"""Pull the agent's tool-call sequence out of the judge prompt.
The judge prompt embeds the trajectory as a JSON array under
`## Agent Trajectory (tool calls and results)`. We don't need a
perfect parse — a one-line summary per tool call (name + a hint
of the arguments) is enough for an audit brief.
"""
marker = "## Agent Trajectory (tool calls and results)"
if marker not in prompt_user:
return []
after = prompt_user.split(marker, 1)[1]
end_markers = [
"## Agent Final Answer",
"## Final Answer",
"## Agent Answer",
"Evaluate across",
]
end_idx = len(after)
for em in end_markers:
i = after.find(em)
if i != -1 and i < end_idx:
end_idx = i
body = after[:end_idx].strip()
# Try to parse as JSON; if that fails, do a permissive name-extract.
try:
traj = json.loads(body)
except Exception:
# Permissive fallback: find every `"name": "..."`.
out = []
i = 0
while True:
i = body.find('"name":', i)
if i == -1:
break
j = body.find('"', i + 7)
k = body.find('"', j + 1)
if j == -1 or k == -1:
break
out.append(body[j + 1 : k])
i = k + 1
return out
out = []
if isinstance(traj, list):
for turn in traj:
for tc in (turn or {}).get("tool_calls", []) or []:
name = tc.get("name", "?")
args = tc.get("arguments", {}) or {}
arg_hint = (
", ".join(f"{k}={args[k]!r}" for k in list(args)[:2])
if isinstance(args, dict)
else ""
)
out.append(f"{name}({arg_hint})" if arg_hint else name)
return out
def _extract_final_answer(prompt_user: str, raw_response: str) -> str:
"""Best-effort extract of the agent's final answer from the prompt."""
for marker in ("## Agent Final Answer", "## Final Answer", "## Agent Answer"):
if marker in prompt_user:
after = prompt_user.split(marker, 1)[1]
end = after.find("Evaluate across")
return (after[:end] if end != -1 else after).strip()[:1500]
return ""
def _row_brief(row: dict) -> dict:
"""Build a compact audit brief for one taxonomy row.
Returns a dict with the source row plus extracted judge-log signals.
Missing judge log -> brief flagged `judge_log_present=False`. Uses
`_exists` / `_read_text` so Windows long paths work the same as POSIX.
"""
log_path = _judge_log_path(row["run_name"], row["scenario_id"], row["trial_index"])
present = _exists(log_path)
out = {
"source_row": row,
"judge_log_path": (
str(log_path.relative_to(REPO_ROOT).as_posix()) if present else None
),
"judge_log_present": present,
"tool_calls": [],
"final_answer": "",
"judge_suggestions": "",
"judge_parsed_dims": {},
}
if not present:
return out
log = json.loads(_read_text(log_path))
out["tool_calls"] = _extract_tool_call_summary(log.get("prompt_user", ""))
out["final_answer"] = _extract_final_answer(
log.get("prompt_user", ""), log.get("raw_response", "")
)
out["judge_suggestions"] = (log.get("parsed_dims") or {}).get("suggestions", "")
out["judge_parsed_dims"] = {
k: v for k, v in (log.get("parsed_dims") or {}).items() if k != "suggestions"
}
return out
def cmd_briefs(args: argparse.Namespace) -> int:
fields, rows = _load_taxonomy_rows()
statuses = (
{args.status} if args.status else {"stratified_sample", "manual_confirmed"}
)
sample = [r for r in rows if r["audit_status"] in statuses]
if args.cell:
sample = [r for r in sample if r["experiment_cell"] == args.cell]
if args.label:
sample = [r for r in sample if r["auto_taxonomy_label"] == args.label]
out_lines: list[str] = []
for r in sample:
b = _row_brief(r)
sr = b["source_row"]
out_lines.append(
f"=== {sr['run_name']} / {sr['scenario_id']} run{int(sr['trial_index']):02d} ==="
)
out_lines.append(
f"cell={sr['experiment_cell']} orch={sr['orchestration_mode']} "
f"model={sr['model_id']} judge={sr['judge_model']}"
)
out_lines.append(
f"auto_taxonomy_label={sr['auto_taxonomy_label']} "
f"failed_dims=({sr['failed_dim_count']}) {sr['failed_dims']}"
)
if not b["judge_log_present"]:
out_lines.append(" [WARN] judge log MISSING at expected path")
out_lines.append("")
continue
out_lines.append(f" judge_log: {b['judge_log_path']}")
out_lines.append(f" parsed_dims: {b['judge_parsed_dims']}")
if b["tool_calls"]:
out_lines.append(
f" tool_calls ({len(b['tool_calls'])}): "
+ ", ".join(b["tool_calls"][:8])
)
if len(b["tool_calls"]) > 8:
out_lines.append(f" ... and {len(b['tool_calls']) - 8} more")
else:
out_lines.append(" tool_calls: (none extracted)")
fa = b["final_answer"].strip().replace("\n", " ")[:240]
if fa:
out_lines.append(f" final_answer: {fa}")
sug = (b["judge_suggestions"] or "").strip().replace("\n", " ")[:300]
if sug:
out_lines.append(f" judge_suggestion: {sug}")
out_lines.append("")
text = "\n".join(out_lines)
if args.out:
Path(args.out).write_text(text, encoding="utf-8")
print(f"wrote {args.out}: {len(sample)} briefs", file=sys.stderr)
else:
sys.stdout.write(text)
return 0
def cmd_add_columns(args: argparse.Namespace) -> int:
"""Merge an audit-decisions sidecar TSV/JSONL into the taxonomy CSV.
Sidecar shape (TSV):
run_name<TAB>scenario_id<TAB>trial_index<TAB>audit_decision
<TAB>berkeley_label<TAB>failure_stage<TAB>audit_note
<TAB>audit_decision_source
Default sidecar path is `data/audit/issue194_decisions.tsv` (committed
so the audit is reproducible from inputs); override with --decisions.
Eligible rows: `audit_status` is `stratified_sample` (initial pass)
or `manual_confirmed` (re-merge from the same sidecar after the
initial flip). Matching keys are `(run_name, scenario_id,
int(trial_index))`. After merge, `audit_status` is set to
`manual_confirmed` (idempotent on re-runs). The other 1,920+ rows
are left alone.
Fails if the sidecar references a key that doesn't match any
eligible row, or if zero rows would be updated — silent no-op
re-runs were a v1 review-finding hazard.
"""
fields, rows = _load_taxonomy_rows()
if any(c not in fields for c in AUDIT_COLUMNS):
# Insert the new columns just after audit_status for grouping.
idx = fields.index("audit_status") + 1
for new_col in reversed(AUDIT_COLUMNS):
if new_col not in fields:
fields.insert(idx, new_col)
for r in rows:
for new_col in AUDIT_COLUMNS:
r.setdefault(new_col, "")
# Load sidecar.
sidecar = Path(args.decisions) if args.decisions else DECISIONS_TSV
decisions: dict[tuple[str, str, int], dict[str, str]] = {}
with sidecar.open(encoding="utf-8", newline="") as f:
reader = csv.DictReader(f, delimiter="\t")
for row in reader:
key = (row["run_name"], row["scenario_id"], int(row["trial_index"]))
decisions[key] = row
eligible_statuses = {"stratified_sample", "manual_confirmed"}
n_updated = 0
matched_keys: set[tuple[str, str, int]] = set()
for r in rows:
if r["audit_status"] not in eligible_statuses:
continue
key = (r["run_name"], r["scenario_id"], int(r["trial_index"]))
d = decisions.get(key)
if not d:
continue
if d["audit_decision"] not in AUDIT_DECISIONS:
raise ValueError(
f"invalid audit_decision {d['audit_decision']!r} for {key}; "
f"must be one of {sorted(AUDIT_DECISIONS)}"
)
if d.get("berkeley_label") and d["berkeley_label"] not in BERKELEY_LABELS:
raise ValueError(
f"invalid berkeley_label {d['berkeley_label']!r} for {key}; "
f"must be one of {sorted(BERKELEY_LABELS)}"
)
if d.get("failure_stage") and d["failure_stage"] not in FAILURE_STAGES:
raise ValueError(
f"invalid failure_stage {d['failure_stage']!r} for {key}; "
f"must be one of {sorted(FAILURE_STAGES)}"
)
src = d.get("audit_decision_source") or ""
if src and not AUDIT_SOURCE_PATTERN.match(src):
raise ValueError(
f"invalid audit_decision_source {src!r} for {key}; expected "
f"`<reviewer-handle>:<YYYY-MM-DD>` (lowercase handle, ISO date)"
)
for col in AUDIT_COLUMNS:
r[col] = d.get(col, "") or ""
r["audit_status"] = "manual_confirmed"
matched_keys.add(key)
n_updated += 1
unmatched = set(decisions.keys()) - matched_keys
if unmatched:
raise ValueError(
f"sidecar has {len(unmatched)} key(s) that don't match any eligible "
f"row in {TAXONOMY_CSV.relative_to(REPO_ROOT)}: "
f"{sorted(unmatched)[:5]}"
)
if n_updated == 0:
raise ValueError(
f"sidecar at {sidecar} updated zero rows; either the sidecar is "
f"empty or no rows are in eligible audit_status "
f"{sorted(eligible_statuses)}."
)
with TAXONOMY_CSV.open("w", encoding="utf-8", newline="") as f:
w = csv.DictWriter(f, fieldnames=fields, lineterminator="\n")
w.writeheader()
for r in rows:
w.writerow(r)
print(
f"merged {n_updated} audit decisions into {TAXONOMY_CSV.relative_to(REPO_ROOT)}",
file=sys.stderr,
)
return 0
def cmd_render(args: argparse.Namespace) -> int:
fields, rows = _load_taxonomy_rows()
audited = [r for r in rows if r["audit_status"] == "manual_confirmed"]
out: list[str] = []
out.append("# Failure taxonomy hand-audit — stratified sample (#194)")
out.append("")
out.append(
f"*Audit pass over the {sum(1 for r in rows if r['audit_status'] in ('manual_confirmed', 'stratified_sample'))}"
"-row stratified sample drawn by `scripts/build_failure_taxonomy.py` "
"from PR #193's `results/metrics/failure_taxonomy_current.csv`. "
f"Audited rows: {len(audited)}.*"
)
out.append("")
out.append(
"Auditor walked each row's judge log under `results/judge_logs/<run_name>"
"/<scenario_id>_run<NN>_judge_log.json`, which embeds the agent's full "
"trajectory + final answer. For each row the auditor decided: "
"(a) `audit_decision` — `confirmed` (auto-label and trajectory symptoms "
"agree), `relabel_suggested` (a concrete better label), or "
"`evidence_thin` (judge symptom too weak/ambiguous to cite); (b) "
"`berkeley_label` — `specification` / `inter_agent_orchestration` / "
"`task_verification` per `docs/failure_taxonomy_evidence.md`; (c) "
"`failure_stage` — earliest stage where the run becomes unrecoverable."
)
out.append("")
# Methodology + Berkeley mapping rule come from the patterns sidecar so
# the rendered doc is reproducible from committed inputs.
patterns_json = (
json.loads(_read_text(PATTERNS_JSON)) if _exists(PATTERNS_JSON) else {}
)
methodology = patterns_json.get("methodology") or {}
if methodology:
out.append(f"## {methodology.get('title', 'Methodology')}")
out.append("")
out.append(methodology.get("body", ""))
out.append("")
supersession = patterns_json.get("current_surface_supersession") or {}
if supersession:
out.append(
f"## {supersession.get('title', 'Current taxonomy surface and supersession')}"
)
out.append("")
out.append(supersession.get("body", ""))
out.append("")
rule = patterns_json.get("berkeley_mapping_rule") or {}
if rule:
out.append(f"## {rule.get('title', 'Berkeley-label mapping rule')}")
out.append("")
out.append(rule.get("body", ""))
out.append("")
stage_rule = patterns_json.get("failure_stage_rule") or {}
if stage_rule:
out.append(f"## {stage_rule.get('title', 'Failure-stage rule')}")
out.append("")
out.append(stage_rule.get("body", ""))
out.append("")
out.append("## Decision counts")
out.append("")
decisions = Counter(r["audit_decision"] for r in audited if r["audit_decision"])
out.append("| audit_decision | count |")
out.append("|---|---:|")
for k in ("confirmed", "relabel_suggested", "evidence_thin"):
out.append(f"| `{k}` | {decisions.get(k, 0)} |")
out.append("")
out.append("## Berkeley taxonomy distribution (audited rows)")
out.append("")
berkeley = Counter(r["berkeley_label"] for r in audited if r["berkeley_label"])
out.append("| berkeley_label | count |")
out.append("|---|---:|")
for k, v in berkeley.most_common():
out.append(f"| `{k}` | {v} |")
out.append("")
out.append("## Failure-stage distribution (audited rows)")
out.append("")
stages = Counter(r["failure_stage"] for r in audited if r["failure_stage"])
out.append("| failure_stage | count |")
out.append("|---|---:|")
for k, v in stages.most_common():
out.append(f"| `{k}` | {v} |")
out.append("")
if patterns_json:
out.append("## Recurring failure patterns")
out.append("")
out.append(
patterns_json.get(
"lead",
"These patterns recur across multiple audited rows and are the "
"strongest paper-citable signals.",
)
)
out.append("")
for i, p in enumerate(patterns_json.get("patterns", []), 1):
cells = ", ".join(p.get("cells", []))
distinct = p.get("distinct_scenarios_count")
distinct_ids = ", ".join(p.get("distinct_scenarios", []))
scope = (
f"{p['count']} rows across {distinct} distinct scenario{'s' if distinct != 1 else ''}"
f" ({distinct_ids})"
if distinct is not None
else f"{p['count']} rows"
)
out.append(
f"{i}. **{p['title']}** — {scope}. {p['description']} "
f"Cells: {cells}. Mitigation candidate: {p['mitigation']}"
)
if "summary_note" in patterns_json:
out.append("")
out.append(patterns_json["summary_note"])
out.append("")
relabels = patterns_json.get("tie_break_relabels") or {}
if relabels.get("rows"):
has_v4 = any("v4" in r for r in relabels["rows"])
if has_v4:
out.append("## Tie-break relabels (v2 → v3 → v4)")
out.append("")
out.append(
"Six rows where v2's first-pass labels needed re-evaluation. "
"v3 changed the Berkeley labels per the latest-irreversible-mistake "
"tie-break rule (correct) but also applied that rule to "
"`failure_stage` (incorrect — stage uses earliest-unrecoverable "
"per `docs/failure_taxonomy_evidence.md:149-151`). v4 keeps "
"the v3 Berkeley relabels and re-evaluates each row's stage "
"independently."
)
out.append("")
out.append(
"| cell | scenario | trial | v2 berkeley / stage | v3 berkeley / stage | v4 berkeley / stage | rationale |"
)
out.append("|---|---|---:|---|---|---|---|")
for r in relabels["rows"]:
v2 = r["v2"]
v3 = r["v3"]
v4 = r.get("v4", v3)
rationale = r["rationale"].replace("|", "\\|")
out.append(
f"| {r['cell']} | {r['scenario_id']} | {r['trial_index']} | "
f"`{v2['berkeley']}` / `{v2['stage']}` | "
f"`{v3['berkeley']}` / `{v3['stage']}` | "
f"`{v4['berkeley']}` / `{v4['stage']}` | {rationale} |"
)
else:
out.append("## Tie-break relabels (v2 → v3)")
out.append("")
out.append(
"Six rows were relabeled in v3 to apply the team's tie-break rule "
"consistently (latest irreversible mistake). v2's labels reflected "
"the auditor's first-pass 'first irreversible mistake' reading; "
"v3 corrects this against `docs/failure_taxonomy_evidence.md:139-145`."
)
out.append("")
out.append(
"| cell | scenario | trial | v2 berkeley / stage | v3 berkeley / stage | rationale |"
)
out.append("|---|---|---:|---|---|---|")
for r in relabels["rows"]:
v2 = r["v2"]
v3 = r["v3"]
rationale = r["rationale"].replace("|", "\\|")
out.append(
f"| {r['cell']} | {r['scenario_id']} | {r['trial_index']} | "
f"`{v2['berkeley']}` / `{v2['stage']}` | "
f"`{v3['berkeley']}` / `{v3['stage']}` | {rationale} |"
)
out.append("")
out.append("## Per-row audit table")
out.append("")
out.append(
"| cell | run | scenario | trial | auto_label | decision | berkeley | stage | note |"
)
out.append("|---|---|---|---:|---|---|---|---|---|")
for r in audited:
run_short = r["run_name"]
if len(run_short) > 60:
run_short = "..." + run_short[-58:]
note = (r["audit_note"] or "").replace("|", "\\|")
out.append(
f"| {r['experiment_cell']} | `{run_short}` | {r['scenario_id']} | "
f"{r['trial_index']} | `{r['auto_taxonomy_label'].removeprefix('low_')}` | "
f"`{r['audit_decision']}` | `{r['berkeley_label']}` | "
f"`{r['failure_stage']}` | {note} |"
)
out.append("")
# Paper-cite candidates: confirmed rows with a clear concrete note.
cited = [
r
for r in audited
if r.get("audit_decision") == "confirmed"
and (r.get("berkeley_label") or "")
and (r.get("failure_stage") or "")
][: args.paper_cite_n]
out.append(f"## Paper-cite candidates (top {len(cited)})")
out.append("")
out.append(
"These rows have a confirmed auto-label, a concrete trajectory-grounded "
"note, and a balanced cell/orchestration/symptom mix. Recommended "
"examples for the paper's failure-analysis section."
)
out.append("")
for i, r in enumerate(cited, 1):
out.append(
f"{i}. **Cell {r['experiment_cell']} / {r['orchestration_mode']} / "
f"{r['scenario_id']} run{int(r['trial_index']):02d}** — "
f"{r['auto_taxonomy_label']} -> `{r['berkeley_label']}` "
f"({r['failure_stage']}). {r['audit_note']} "
f"Source: `{r['trajectory_file']}`."
)
out.append("")
text = "\n".join(out)
if args.out:
Path(args.out).write_text(text, encoding="utf-8")
print(
f"wrote {args.out}: {len(audited)} audited rows, {len(cited)} paper-cite candidates",
file=sys.stderr,
)
else:
sys.stdout.write(text)
return 0
def main(argv: list[str] | None = None) -> int:
p = argparse.ArgumentParser(description=__doc__)
sp = p.add_subparsers(dest="cmd", required=True)
pb = sp.add_parser("briefs", help="emit per-row briefs for the auditor")
pb.add_argument("--out", type=str, default=None)
pb.add_argument("--cell", type=str, default=None, help="filter by experiment_cell")
pb.add_argument(
"--label", type=str, default=None, help="filter by auto_taxonomy_label"
)
pb.add_argument(
"--status",
type=str,
default=None,
choices=("stratified_sample", "manual_confirmed"),
help="filter by audit_status (default: include both)",
)
pb.set_defaults(fn=cmd_briefs)
pa = sp.add_parser(
"add-columns",
help="merge a hand-edited audit-decisions sidecar TSV into the taxonomy CSV",
)
pa.add_argument(
"--decisions",
type=str,
default=None,
help=(
"path to the audit-decisions TSV; defaults to "
"data/audit/issue194_decisions.tsv (committed)"
),
)
pa.set_defaults(fn=cmd_add_columns)
pr = sp.add_parser(
"render", help="render markdown audit summary from the augmented CSV"
)
pr.add_argument("--out", type=str, default=None)
pr.add_argument(
"--paper-cite-n",
type=int,
default=10,
help="max paper-cite candidates to surface",
)
pr.set_defaults(fn=cmd_render)
args = p.parse_args(argv)
return args.fn(args)
if __name__ == "__main__":
raise SystemExit(main())