russell@unturf.com
d3b78025ff
New bench/teacher_judge.py reads judge_input.jsonl, cross-judges
each (question, model, answer) using a DIFFERENT model than the
answerer (default: hermes-answers → qwen judges, qwen-answers →
hermes judges). Writes verdicts.jsonl + markdown summary including
a false-STRICT highlight section.
Key implementation details:
- Reuses _judge_prompt from cross_model_selfplay so iterating on
the prompt template doesn't require re-running benches.
- Sends chat_template_kwargs:{enable_thinking:false} on every call
— required for qwen.ai.unturf.com (llama.cpp deepseek-reasoning
format) where reasoning eats max_tokens before producing content.
vLLM (hermes) silently ignores the unknown kwarg.
- Prompt template tightened: previously said "Reply in format
VERDICT: <rationale>" which qwen took literally, replying with
the word VERDICT instead of the verdict token. Now explicit:
"write the chosen verdict word itself".
Live cross-judge result on 151-row 76-question bench:
hermes (judged by qwen): 44 CORRECT, 23 WRONG, 9 PARTIAL → 58% accuracy
qwen (judged by hermes): 45 CORRECT, 6 WRONG, 11 PARTIAL → 73% accuracy
Qwen is materially more accurate despite costing 1.8× more — matches
fox's "qwen slightly outperforms" intuition with hard numbers.
8 false-STRICTs surfaced including Q12 (hermes conflated Roman
Empire with HRE), Q48 (hermes answered with Niger River info on a
Nile question), and Q72 (both models missed the Game of Thrones
reference in "winter is coming").
520 lines
20 KiB
Python
520 lines
20 KiB
Python
"""Cross-model self-play bench — same question, multiple models, compare.
|
||
|
||
The pattern: the uncloseai-cli harness can call arborist twice per
|
||
question with different models to get two opinions. This script bakes
|
||
that pattern in as a benchmark. For each question in a fixture, run
|
||
arborist once per configured model, then tabulate:
|
||
|
||
* audit_mode per model (EVIDENCE-WARRANTED → ANCHOR-WARRANTED →
|
||
POINTER-LINKED → UNGROUNDED; HYBRID variants flatten to base)
|
||
* agreement on primary source URI
|
||
* grounding rate per model
|
||
* implied spend in fox's natural unit: $/1000 grounded answers
|
||
(Hermes $0.09, Qwen $0.16 by default — override via --baseline)
|
||
* latency per call
|
||
|
||
The COGS unit is **dollars per 1000 grounded answers**, an outcome
|
||
metric — failed/ungrounded calls are already amortized into the
|
||
baseline. A model's baseline × grounded_count / 1000 = implied spend
|
||
for those grounded answers.
|
||
|
||
Outputs:
|
||
|
||
bench/cross_model_results/<utc-iso>.jsonl one row per (question, model)
|
||
bench/cross_model_results/<utc-iso>.md markdown summary
|
||
|
||
Usage:
|
||
|
||
make bench-cross-model # smoke fixture, Hermes + Qwen
|
||
make bench-cross-model BENCH_CM_QUESTIONS=bench/qa_questions.txt
|
||
|
||
# custom baseline (dollars per 1000 grounded answers):
|
||
python3 bench/cross_model_selfplay.py --baseline 'hermes=0.09,qwen=0.16'
|
||
|
||
# custom model list (name=endpoint:model_id, comma-separated):
|
||
python3 bench/cross_model_selfplay.py \\
|
||
--models 'hermes=https://hermes.ai.unturf.com/v1:adamo1139/Hermes-3-Llama-3.1-8B-FP8-Dynamic,qwen=https://qwen.ai.unturf.com/v1:Qwen3.6-27B-UD-Q4_K_XL.gguf'
|
||
|
||
# emit teacher-judge prompts for downstream correctness grading:
|
||
python3 bench/cross_model_selfplay.py --teacher-model-judge
|
||
# → produces <ts>.judge_input.jsonl with one row per (question, model);
|
||
# feed to a teacher (claude -p, remote API, etc.) to detect false STRICTs
|
||
# — answers that audit STRICT but are factually wrong.
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import argparse
|
||
import datetime as _dt
|
||
import json
|
||
import os
|
||
import subprocess
|
||
import sys
|
||
import time
|
||
from collections import defaultdict
|
||
from pathlib import Path
|
||
|
||
REPO = Path(__file__).resolve().parent.parent
|
||
ARBORIST = REPO / ".venv" / "bin" / "arborist"
|
||
|
||
DEFAULT_MODELS = [
|
||
("hermes", "https://hermes.ai.unturf.com/v1",
|
||
"adamo1139/Hermes-3-Llama-3.1-8B-FP8-Dynamic"),
|
||
("qwen", "https://qwen.ai.unturf.com/v1",
|
||
"Qwen3.6-27B-UD-Q4_K_XL.gguf"),
|
||
]
|
||
# Per fox 2026-05-31 — baseline COGS in dollars per 1000 grounded
|
||
# answers. Hermes (8B) on 3090, Qwen (30B) on 4090. The baseline
|
||
# already amortizes ungrounded calls — it's an outcome metric, not a
|
||
# per-call price. Override via --baseline.
|
||
DEFAULT_DOLLARS_PER_1K_GROUNDED = {"hermes": 0.09, "qwen": 0.16}
|
||
|
||
# audit_mode → numeric rank for "stronger" comparison. Strips
|
||
# "-PARTIAL" suffix that HYBRID modes carry.
|
||
_AUDIT_RANK = {
|
||
"EVIDENCE-WARRANTED": 3,
|
||
"ANCHOR-WARRANTED": 2,
|
||
"POINTER-LINKED": 1,
|
||
"UNGROUNDED": 0,
|
||
"STRICT": 2, # legacy alias for ANCHOR-WARRANTED
|
||
"HYBRID": 1, # legacy alias for POINTER-LINKED
|
||
}
|
||
|
||
|
||
def _audit_rank(mode: str | None) -> int:
|
||
if not mode:
|
||
return 0
|
||
base = mode.replace("-PARTIAL", "")
|
||
return _AUDIT_RANK.get(base, 0)
|
||
|
||
|
||
def _is_grounded(mode: str | None) -> bool:
|
||
"""Anything not UNGROUNDED counts as grounded for cost accounting."""
|
||
if not mode:
|
||
return False
|
||
return not mode.startswith("UNGROUNDED")
|
||
|
||
|
||
def _read_questions(path: Path) -> list[str]:
|
||
out: list[str] = []
|
||
for line in path.read_text().splitlines():
|
||
line = line.strip()
|
||
if not line or line.startswith("#"):
|
||
continue
|
||
out.append(line)
|
||
return out
|
||
|
||
|
||
def _parse_models(spec: str) -> list[tuple[str, str, str]]:
|
||
"""Parse name=endpoint:model_id,... into a list of tuples."""
|
||
out = []
|
||
for chunk in spec.split(","):
|
||
chunk = chunk.strip()
|
||
if not chunk:
|
||
continue
|
||
if "=" not in chunk or ":" not in chunk.split("=", 1)[1]:
|
||
raise SystemExit(
|
||
f"--models entry must be name=endpoint:model_id, got: {chunk!r}"
|
||
)
|
||
name, rest = chunk.split("=", 1)
|
||
# split on the LAST colon — model_id may contain colons (rare)
|
||
# but endpoint always ends at the path boundary, so split on
|
||
# the colon BEFORE a non-slash. Simpler: assume model_id
|
||
# follows the first colon AFTER the endpoint scheme://host[:port]/path.
|
||
# We split on " :" sentinel — but easiest: rsplit once on `:`
|
||
# because vLLM model_ids don't contain colons in practice.
|
||
endpoint, model_id = rest.rsplit(":", 1)
|
||
out.append((name.strip(), endpoint.strip(), model_id.strip()))
|
||
return out
|
||
|
||
|
||
def _parse_baseline(spec: str | None) -> dict[str, float]:
|
||
"""Parse name=dollars_per_1k_grounded,... into a dict."""
|
||
if not spec:
|
||
return dict(DEFAULT_DOLLARS_PER_1K_GROUNDED)
|
||
out = dict(DEFAULT_DOLLARS_PER_1K_GROUNDED)
|
||
for chunk in spec.split(","):
|
||
if "=" not in chunk:
|
||
continue
|
||
name, dollars = chunk.split("=", 1)
|
||
try:
|
||
out[name.strip()] = float(dollars.strip())
|
||
except ValueError:
|
||
pass
|
||
return out
|
||
|
||
|
||
def _run_one(
|
||
question: str, *, shards_dir: Path | None, endpoint: str, model: str,
|
||
top_k: int, timeout_s: int, burn: bool,
|
||
) -> dict:
|
||
cmd = [str(ARBORIST)]
|
||
if shards_dir:
|
||
cmd += ["--shards-dir", str(shards_dir)]
|
||
cmd += [
|
||
"query", "--json",
|
||
"--top-k", str(top_k),
|
||
"--endpoint", endpoint,
|
||
"--model", model,
|
||
]
|
||
if burn:
|
||
cmd.append("--burn")
|
||
cmd.append(question)
|
||
t0 = time.time()
|
||
try:
|
||
proc = subprocess.run(
|
||
cmd, capture_output=True, text=True, timeout=timeout_s,
|
||
env={**os.environ, "ARBORIST_PROGRESS": "0"},
|
||
)
|
||
except subprocess.TimeoutExpired:
|
||
return {"_error": f"timeout after {timeout_s}s",
|
||
"_elapsed_s": float(timeout_s)}
|
||
elapsed = round(time.time() - t0, 2)
|
||
if proc.returncode not in (0, 1):
|
||
# Exit 1 is "ungrounded" — still has JSON. Other codes are real errors.
|
||
return {"_error": f"exit {proc.returncode}",
|
||
"_stderr": (proc.stderr or "")[:500],
|
||
"_elapsed_s": elapsed}
|
||
try:
|
||
data = json.loads(proc.stdout)
|
||
except json.JSONDecodeError as e:
|
||
return {"_error": f"parse: {e}",
|
||
"_stdout_head": proc.stdout[:200],
|
||
"_elapsed_s": elapsed}
|
||
data["_elapsed_s"] = elapsed
|
||
data["_exit_code"] = proc.returncode
|
||
return data
|
||
|
||
|
||
def _primary_uri(result: dict) -> str:
|
||
srcs = result.get("sources") or []
|
||
primary = next(
|
||
(s for s in srcs if s.get("source_role") == "primary_answer_source"),
|
||
srcs[0] if srcs else {},
|
||
)
|
||
return primary.get("document_uri") or ""
|
||
|
||
|
||
def _stronger_model(per_model: dict[str, dict]) -> str:
|
||
"""Return the name of the model with the higher audit_mode rank
|
||
(ties broken by alpha for stability). Empty string if all ungrounded."""
|
||
best = ""
|
||
best_rank = -1
|
||
for name, res in sorted(per_model.items()):
|
||
r = _audit_rank(res.get("audit_mode"))
|
||
if r > best_rank:
|
||
best, best_rank = name, r
|
||
return best if best_rank > 0 else ""
|
||
|
||
|
||
def _judge_prompt(question: str, model_name: str, audit_mode: str, answer: str) -> str:
|
||
"""Build the prompt a teacher model receives to grade one (question, answer)
|
||
pair. Asks for a single verdict token + a one-sentence rationale.
|
||
|
||
Designed to be model-agnostic — any teacher (Opus via `claude -p`, a remote
|
||
API, a Hermes/Qwen self-judge run) can consume it. The teacher should reply
|
||
starting with one of CORRECT/WRONG/PARTIAL/UNCERTAIN followed by a colon
|
||
and a one-sentence rationale.
|
||
"""
|
||
return (
|
||
"You are a hermetic judge grading the factual correctness of one answer "
|
||
"to one question. Use your own parametric knowledge as ground truth. "
|
||
"Ignore the answer's stylistic quality and whether it cites sources — "
|
||
"judge only whether the factual claim is true.\n\n"
|
||
f"Question: {question}\n\n"
|
||
f"Answer: {answer}\n\n"
|
||
"Reply with exactly one word from the set "
|
||
"{CORRECT, WRONG, PARTIAL, UNCERTAIN}, followed by a colon and a "
|
||
"one-sentence rationale. Do not write the literal word VERDICT — "
|
||
"write the chosen verdict word itself.\n\n"
|
||
"Meanings:\n"
|
||
" CORRECT — the answer's main factual claim is true\n"
|
||
" WRONG — the answer's main factual claim is false\n"
|
||
" PARTIAL — main claim correct but missing required detail "
|
||
"(or correct on part, wrong on part)\n"
|
||
" UNCERTAIN — you don't know the domain well enough to judge\n\n"
|
||
"Examples of valid replies:\n"
|
||
" CORRECT: Paris is the capital of France.\n"
|
||
" WRONG: The Nile begins at Lake Victoria, not the Niger Highlands.\n"
|
||
" PARTIAL: Names some members but misses Maggie Simpson.\n"
|
||
" UNCERTAIN: I do not know this obscure regional fact.\n"
|
||
)
|
||
|
||
|
||
def _emit_judge_inputs(
|
||
rows: list[dict], out_path: Path,
|
||
) -> int:
|
||
"""Write one JSONL row per (question, model) ready for a teacher to grade.
|
||
Skips rows that errored or returned no answer text. Returns count written."""
|
||
written = 0
|
||
with out_path.open("w") as f:
|
||
for r in rows:
|
||
res = r.get("result") or {}
|
||
if "_error" in res:
|
||
continue
|
||
answer = (res.get("answer_text") or "").strip()
|
||
if not answer:
|
||
continue
|
||
audit = res.get("audit_mode") or "UNKNOWN"
|
||
payload = {
|
||
"question": r["question"],
|
||
"model_name": r["model_name"],
|
||
"audit_mode": audit,
|
||
"answer_text": answer,
|
||
"judge_prompt": _judge_prompt(
|
||
r["question"], r["model_name"], audit, answer,
|
||
),
|
||
}
|
||
f.write(json.dumps(payload, ensure_ascii=False) + "\n")
|
||
written += 1
|
||
return written
|
||
|
||
|
||
def _summarize(
|
||
rows: list[dict], models: list[tuple[str, str, str]],
|
||
baseline_per_1k: dict[str, float],
|
||
) -> str:
|
||
"""Build the markdown summary. COGS unit: $ per 1k grounded answers."""
|
||
by_q: dict[str, dict[str, dict]] = defaultdict(dict)
|
||
for r in rows:
|
||
by_q[r["question"]][r["model_name"]] = r["result"]
|
||
|
||
n_q = len(by_q)
|
||
model_names = [m[0] for m in models]
|
||
out = []
|
||
out.append(f"# Cross-model self-play bench — {_dt.datetime.now(_dt.timezone.utc).isoformat()}")
|
||
out.append("")
|
||
out.append(f"Questions: {n_q}. Models: {', '.join(model_names)}.")
|
||
out.append("")
|
||
out.append("COGS unit: **dollars per 1000 grounded answers** (fox's natural unit). "
|
||
"Implied spend uses each model's baseline × grounded_count / 1000.")
|
||
out.append("")
|
||
|
||
# Per-model summary
|
||
out.append("## Per-model summary")
|
||
out.append("")
|
||
out.append("| model | grounded | rate | avg latency | baseline $/1k grounded | implied spend |")
|
||
out.append("|---|---:|---:|---:|---:|---:|")
|
||
for name in model_names:
|
||
grounded = sum(
|
||
1 for q in by_q if _is_grounded(by_q[q].get(name, {}).get("audit_mode"))
|
||
)
|
||
total = sum(1 for q in by_q if name in by_q[q] and "_error" not in by_q[q][name])
|
||
rate = (grounded / total) if total else 0.0
|
||
avg_lat = (
|
||
sum(by_q[q].get(name, {}).get("_elapsed_s", 0) for q in by_q) / max(total, 1)
|
||
)
|
||
base = baseline_per_1k.get(name, 0.0)
|
||
spend = grounded * base / 1000
|
||
out.append(
|
||
f"| {name} | {grounded}/{total} | {rate:.0%} "
|
||
f"| {avg_lat:.1f}s | ${base:.2f} | ${spend:.6f} |"
|
||
)
|
||
out.append("")
|
||
|
||
# Agreement
|
||
out.append("## Per-question results")
|
||
out.append("")
|
||
header = ["question"] + model_names + ["stronger", "primary agree"]
|
||
out.append("| " + " | ".join(header) + " |")
|
||
out.append("|" + "|".join(["---"] * len(header)) + "|")
|
||
for q in by_q:
|
||
cells = [q[:64] + ("…" if len(q) > 64 else "")]
|
||
uris = set()
|
||
for name in model_names:
|
||
res = by_q[q].get(name, {})
|
||
audit = res.get("audit_mode") or "—"
|
||
if "_error" in res:
|
||
audit = f"ERR ({res['_error'][:30]})"
|
||
cells.append(audit)
|
||
uri = _primary_uri(res)
|
||
if uri:
|
||
uris.add(uri)
|
||
cells.append(_stronger_model(by_q[q]) or "—")
|
||
cells.append("✓" if len(uris) == 1 else ("✗" if len(uris) > 1 else "—"))
|
||
out.append("| " + " | ".join(cells) + " |")
|
||
out.append("")
|
||
|
||
# Cost analysis
|
||
out.append("## Cost analysis (north star: $/1k grounded)")
|
||
out.append("")
|
||
total_spend = 0.0
|
||
total_grounded = 0
|
||
for name in model_names:
|
||
base = baseline_per_1k.get(name, 0.0)
|
||
n_g = sum(1 for q in by_q if _is_grounded(by_q[q].get(name, {}).get("audit_mode")))
|
||
spent = n_g * base / 1000
|
||
total_spend += spent
|
||
total_grounded += n_g
|
||
out.append(f"* **{name}** — {n_g} grounded × ${base:.2f}/1k = ${spent:.6f}")
|
||
out.append("")
|
||
out.append(f"**Total implied spend**: ${total_spend:.6f}. **Total grounded answers**: {total_grounded}.")
|
||
if total_grounded:
|
||
# Effective $/1k = total_spend × (1000 / total_grounded)
|
||
out.append(f"**Effective COGS**: ${total_spend * 1000 / total_grounded:.4f} per 1k grounded "
|
||
f"(weighted across models in this run).")
|
||
out.append("")
|
||
|
||
# Routing hint — cheap-first cascade
|
||
out.append("## Cheap-first cascade (try cheapest, escalate on UNGROUNDED)")
|
||
out.append("")
|
||
ordered = sorted(model_names, key=lambda n: baseline_per_1k.get(n, 0))
|
||
# Per-q: which model ultimately grounded? Count by source.
|
||
cascade_grounded_by_model: dict[str, int] = defaultdict(int)
|
||
cascade_total_grounded = 0
|
||
for q in by_q:
|
||
for name in ordered:
|
||
res = by_q[q].get(name, {})
|
||
if "_error" in res:
|
||
continue
|
||
if _is_grounded(res.get("audit_mode")):
|
||
cascade_grounded_by_model[name] += 1
|
||
cascade_total_grounded += 1
|
||
break
|
||
# else fall through to next model in cascade
|
||
cascade_spend = sum(
|
||
cnt * baseline_per_1k.get(name, 0) / 1000
|
||
for name, cnt in cascade_grounded_by_model.items()
|
||
)
|
||
if cascade_total_grounded:
|
||
breakdown = ", ".join(
|
||
f"{cnt} from {name}" for name, cnt in cascade_grounded_by_model.items()
|
||
)
|
||
cascade_cogs = cascade_spend * 1000 / cascade_total_grounded
|
||
out.append(
|
||
f"Cascade ({' → '.join(ordered)}): {cascade_total_grounded}/{n_q} grounded "
|
||
f"({breakdown}). Implied spend ${cascade_spend:.6f} → "
|
||
f"**${cascade_cogs:.4f}/1k grounded**."
|
||
)
|
||
# Compare to always-most-expensive (defensive baseline)
|
||
most_expensive = ordered[-1]
|
||
max_base = baseline_per_1k.get(most_expensive, 0)
|
||
always_grounded = sum(
|
||
1 for q in by_q
|
||
if _is_grounded(by_q[q].get(most_expensive, {}).get("audit_mode"))
|
||
)
|
||
always_spend = always_grounded * max_base / 1000
|
||
if always_grounded:
|
||
always_cogs = always_spend * 1000 / always_grounded
|
||
delta_pct = (cascade_cogs / always_cogs - 1) * 100 if always_cogs else 0
|
||
sign = "+" if delta_pct >= 0 else ""
|
||
out.append("")
|
||
out.append(
|
||
f"vs always-{most_expensive}: {always_grounded}/{n_q} grounded, "
|
||
f"${always_spend:.6f} spend, ${always_cogs:.4f}/1k. "
|
||
f"Cascade delta: **{sign}{delta_pct:.1f}%**."
|
||
)
|
||
out.append("")
|
||
out.append("Caveat: cascade math credits each grounded answer to its source model's baseline. "
|
||
"Real-world failed-call overhead (an UNGROUNDED hermes call that escalates to qwen) "
|
||
"is not separately accounted — it's amortized into hermes's baseline by definition.")
|
||
out.append("")
|
||
return "\n".join(out)
|
||
|
||
|
||
def main() -> int:
|
||
ap = argparse.ArgumentParser(description=__doc__)
|
||
ap.add_argument(
|
||
"--questions", type=Path,
|
||
default=Path("bench/qa_questions_smoke.txt"),
|
||
)
|
||
ap.add_argument("--shards-dir", type=Path, default=None)
|
||
ap.add_argument(
|
||
"--models", default=None,
|
||
help="comma-separated name=endpoint:model_id; defaults to hermes + qwen",
|
||
)
|
||
ap.add_argument(
|
||
"--baseline", default=None,
|
||
help="comma-separated name=dollars per 1k grounded answers; "
|
||
"defaults to hermes=0.09,qwen=0.16 (fox's COGS unit)",
|
||
)
|
||
ap.add_argument("--top-k", type=int, default=8)
|
||
# 600s default — queue depth on shared endpoints (e.g. hermes
|
||
# under contention from another agent) can push a single call
|
||
# past 3 minutes even when the model is healthy. 180s was too
|
||
# aggressive against real-world infrastructure load.
|
||
ap.add_argument("--timeout", type=int, default=600)
|
||
ap.add_argument(
|
||
"--burn", action="store_true",
|
||
help="force fresh inference (skip cache) on every call",
|
||
)
|
||
ap.add_argument("--limit", type=int, default=0, help="cap questions (0 = no cap)")
|
||
ap.add_argument(
|
||
"--out-dir", type=Path, default=Path("bench/cross_model_results"),
|
||
)
|
||
ap.add_argument(
|
||
"--teacher-model-judge", action="store_true",
|
||
help="emit <ts>.judge_input.jsonl alongside results — one row per "
|
||
"(question, model) with a teacher-ready prompt for grading. "
|
||
"Feed to any teacher (claude -p, remote API, Hermes self-judge) "
|
||
"to get correctness verdicts. STRICT audit ≠ correct — this is "
|
||
"how you measure false-STRICTs.",
|
||
)
|
||
args = ap.parse_args()
|
||
|
||
questions = _read_questions(args.questions)
|
||
if args.limit:
|
||
questions = questions[: args.limit]
|
||
if not questions:
|
||
print(f"no questions in {args.questions}", file=sys.stderr)
|
||
return 1
|
||
|
||
if args.models:
|
||
models = _parse_models(args.models)
|
||
else:
|
||
models = list(DEFAULT_MODELS)
|
||
baseline = _parse_baseline(args.baseline)
|
||
|
||
if not ARBORIST.exists():
|
||
print(f"arborist binary not found: {ARBORIST}", file=sys.stderr)
|
||
return 2
|
||
|
||
args.out_dir.mkdir(parents=True, exist_ok=True)
|
||
ts = _dt.datetime.now(_dt.timezone.utc).strftime("%Y%m%dT%H%M%SZ")
|
||
jsonl_path = args.out_dir / f"{ts}.jsonl"
|
||
md_path = args.out_dir / f"{ts}.md"
|
||
|
||
rows: list[dict] = []
|
||
print(f"# {len(questions)} questions × {len(models)} models = "
|
||
f"{len(questions) * len(models)} calls", file=sys.stderr)
|
||
with jsonl_path.open("w") as f:
|
||
for q in questions:
|
||
for name, endpoint, model_id in models:
|
||
print(f" [{name}] {q[:80]}", file=sys.stderr)
|
||
res = _run_one(
|
||
q, shards_dir=args.shards_dir,
|
||
endpoint=endpoint, model=model_id,
|
||
top_k=args.top_k, timeout_s=args.timeout, burn=args.burn,
|
||
)
|
||
row = {
|
||
"question": q,
|
||
"model_name": name,
|
||
"model_id": model_id,
|
||
"endpoint": endpoint,
|
||
"result": res,
|
||
}
|
||
rows.append(row)
|
||
f.write(json.dumps(row, ensure_ascii=False) + "\n")
|
||
f.flush()
|
||
audit = res.get("audit_mode") or res.get("_error") or "?"
|
||
elapsed = res.get("_elapsed_s", "?")
|
||
print(f" → {audit} ({elapsed}s)", file=sys.stderr)
|
||
|
||
md_path.write_text(_summarize(rows, models, baseline))
|
||
print(f"\nJSONL: {jsonl_path}", file=sys.stderr)
|
||
print(f"Summary: {md_path}", file=sys.stderr)
|
||
|
||
if args.teacher_model_judge:
|
||
judge_input_path = args.out_dir / f"{ts}.judge_input.jsonl"
|
||
n = _emit_judge_inputs(rows, judge_input_path)
|
||
print(f"Teacher-judge inputs: {judge_input_path} ({n} rows)", file=sys.stderr)
|
||
print(f" Feed to your teacher of choice (e.g. `claude -p` per row) "
|
||
f"to grade. Verdicts are CORRECT|WRONG|PARTIAL|UNCERTAIN.",
|
||
file=sys.stderr)
|
||
|
||
print()
|
||
print(md_path.read_text())
|
||
return 0
|
||
|
||
|
||
if __name__ == "__main__":
|
||
sys.exit(main())
|