File size: 28,658 Bytes
43d29b8 8c42697 43d29b8 8c42697 43d29b8 f7d51cb 43d29b8 f7d51cb 43d29b8 f7d51cb 43d29b8 f7d51cb 43d29b8 f7d51cb 43d29b8 2df3df7 2a58472 2df3df7 8c42697 2df3df7 8c42697 2df3df7 8c42697 2df3df7 8c42697 2df3df7 43d29b8 2df3df7 43d29b8 2df3df7 2a58472 43d29b8 f7d51cb 43d29b8 2df3df7 43d29b8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 | #!/usr/bin/env python3
# =============================================================================
# run_experiments.py
# -----------------------------------------------------------------------------
# Responsible for: The single entrypoint that runs every claim's experiment and
# writes results to an output directory as JSON + CSV.
# Role in project: Runs identically on a laptop (--stage smoke, mock backend)
# and on an HF GPU Job (real vLLM backend). One script, one
# code path, so the smoke test actually exercises the real thing.
# Assumptions: Reads backend config from EDUMIRROR_* env vars (see llm.py).
# Writes to --out. Never reads the network except via the LLM.
# =============================================================================
"""Experiment driver for the EduMirror reproduction.
Stages:
smoke -- tiny mock-backed run proving the pipeline end to end
claim2 -- kindergarten scalability at 5/15/30 agents (Table 1)
claim3 -- dual measurement: bullying dynamics + RSES construct validity
claim4 -- pairwise win-rate heatmap across scenarios (Figure 4)
claim5 -- election intervention strategies (Figure 7)
all -- claim2 + claim3 + claim4 + claim5
Every stage writes <out>/<stage>.json and appends a row-per-record CSV so the
logbook can attach both the raw traces and the tables.
"""
from __future__ import annotations
import argparse
import csv
import json
import random
import statistics
import sys
import time
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from edumirror.agents import AGENT_REGISTRY
from edumirror.gm import run_episode
from edumirror.judgecheck import CORRUPTIONS, probe_absolute, probe_pairwise
from edumirror.llm import llm_from_env
from edumirror.measure import (
RATING_METRICS,
pairwise_compare,
rate_behaviour,
rate_competition,
survey_rses,
survey_svo,
win_rate,
)
from edumirror.needs import init_need_state_from_profile
from edumirror.scenarios import (
BULLYING_INTERVENTIONS,
ELECTION_INTERVENTIONS,
bullying_scenario,
class_task_scenario,
election_scenario,
kindergarten_scenario,
representative_scenarios,
study_group_scenario,
)
#: Methods compared, in the paper's Table 1 column order.
METHODS = ["EduMirror", "LLMob", "BabyAGI", "D2A", "ReAct"]
#: Corruption names, for filtering probe output when logging.
CORRUPTION_KEYS = set(CORRUPTIONS)
def log(msg: str) -> None:
"""Print a timestamped progress line, flushed for live HF Job logs."""
print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)
def write_json(path: Path, obj) -> None:
"""Write `obj` as indented JSON, creating parent directories."""
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(obj, indent=2, default=str))
def write_csv(path: Path, rows: list[dict]) -> None:
"""Write a list of flat dicts as CSV (no-op when empty)."""
if not rows:
return
path.parent.mkdir(parents=True, exist_ok=True)
keys = list(rows[0])
with path.open("w", newline="") as fh:
writer = csv.DictWriter(fh, fieldnames=keys)
writer.writeheader()
writer.writerows(rows)
# -----------------------------------------------------------------------------
# Claim 2: kindergarten scalability (Table 1).
# -----------------------------------------------------------------------------
def stage_claim2(out: Path, args) -> dict:
"""
Run the kindergarten scalability experiment across group sizes and methods.
Args:
out: Output directory.
args: Parsed CLI args (group_sizes, n_steps, n_seeds, concurrency).
Returns:
A results dict: {"table": {size: {method: avg}}, "records": [...]}.
Why this design:
Every (size, method, seed) cell runs the SAME scenario object and the
SAME seed, so the only difference across methods is architecture. We run
n_seeds independent episodes per cell and report the mean, because a
single LLM episode is high-variance and a one-shot comparison would be
indistinguishable from noise. The paper does not state its own n, so we
report ours explicitly rather than matching an unknown.
"""
agent_llm = llm_from_env("agent")
judge = llm_from_env("judge")
records: list[dict] = []
for size in args.group_sizes:
scenario = kindergarten_scenario(n_agents=size, n_steps=args.n_steps)
for method in METHODS:
for seed in range(args.n_seeds):
t0 = time.time()
log(f"claim2: size={size} method={method} seed={seed} ...")
try:
ep = run_episode(
scenario, method, agent_llm, seed=seed, concurrency=args.concurrency
)
except Exception as exc: # noqa: BLE001
# One failed episode must not lose the other 44 cells.
log(f" EPISODE FAILED: {exc}")
records.append(
{"size": size, "method": method, "seed": seed, "error": str(exc)}
)
continue
rating = rate_behaviour(judge, ep.behaviour_text(), scenario.setting)
row = {
"size": size,
"method": method,
"seed": seed,
"elapsed_s": round(time.time() - t0, 1),
**{m: rating.scores.get(m) for m in RATING_METRICS},
"average": rating.average(list(RATING_METRICS)),
"rationale": rating.rationale,
}
records.append(row)
log(f" -> avg={row['average']} ({row['elapsed_s']}s)")
ep.save(out / "episodes" / f"kindergarten_{size}_{method}_{seed}.json")
# Aggregate to the Table 1 shape: mean of the per-seed averages.
table: dict[str, dict[str, float]] = {}
for size in args.group_sizes:
table[str(size)] = {}
for method in METHODS:
vals = [
r["average"]
for r in records
if r.get("size") == size
and r.get("method") == method
and isinstance(r.get("average"), (int, float))
]
table[str(size)][method] = round(statistics.mean(vals), 3) if vals else None
results = {"table": table, "records": records, "n_seeds": args.n_seeds,
"n_steps": args.n_steps}
write_json(out / "claim2.json", results)
write_csv(out / "claim2_records.csv", [r for r in records if "error" not in r])
log(f"claim2 table: {json.dumps(table)}")
return results
# -----------------------------------------------------------------------------
# Claim 3: dual measurement protocol.
# -----------------------------------------------------------------------------
def stage_claim3(out: Path, args) -> dict:
"""
Run the bullying case study with dual-track measurement + RSES validity check.
Args:
out: Output directory.
args: Parsed CLI args.
Returns:
Results dict with per-arm need dynamics, RSES pre/post, and the
internal-vs-external correlation.
Why the pre/post RSES design:
Appendix F.5 administers the RSES to Alice before and after the bullying
incidents and compares delta-RSES (external instrument) against
delta-Value on the "self worth"/"sense of respect" dimensions (internal
state). Agreement between them is the construct-validity evidence for
Claim 3: it shows the internal value numbers mean what they claim to.
We reproduce exactly that comparison.
"""
agent_llm = llm_from_env("agent")
surveyor = llm_from_env("judge")
records: list[dict] = []
for arm in BULLYING_INTERVENTIONS:
for seed in range(args.n_seeds):
scenario = bullying_scenario(arm, n_steps=args.n_steps)
# Pin Alice's initial state so arms are compared under "identical
# initial settings" (Sec 4.3).
need_states = {
"Alice": init_need_state_from_profile(
random.Random(seed), baseline_current=6.0
)
}
pre_state = {
"category_means": need_states["Alice"].category_means(),
"needs_current": dict(need_states["Alice"].current),
}
pre = survey_rses(surveyor, pre_state, "Alice, a 14-year-old student")
log(f"claim3: arm={arm} seed={seed} ...")
try:
ep = run_episode(
scenario, "EduMirror", agent_llm, seed=seed,
need_states=need_states, concurrency=args.concurrency,
)
except Exception as exc: # noqa: BLE001
log(f" EPISODE FAILED: {exc}")
continue
alice = ep.internal_states.get("Alice", {})
post = survey_rses(surveyor, alice, "Alice, a 14-year-old student")
# Internal metric: mean of the two self-esteem sub-dimensions
# (Appendix F.5's delta-Value definition).
esteem_dims = ("self worth", "sense of respect")
pre_val = statistics.mean(pre_state["needs_current"][d] for d in esteem_dims)
post_val = statistics.mean(
alice.get("needs_current", pre_state["needs_current"])[d] for d in esteem_dims
)
records.append({
"arm": arm,
"seed": seed,
"rses_pre": pre["total"],
"rses_post": post["total"],
"rses_valid": pre["valid"] and post["valid"],
"delta_rses": (post["total"] - pre["total"])
if (pre["valid"] and post["valid"]) else None,
"value_pre": round(pre_val, 3),
"value_post": round(post_val, 3),
"delta_value": round(post_val - pre_val, 3),
**{f"cat_{k}": round(v, 3)
for k, v in alice.get("category_means", {}).items()},
})
ep.save(out / "episodes" / f"bullying_{arm}_{seed}.json")
# Construct validity: does the external instrument track the internal state?
paired = [
(r["delta_value"], r["delta_rses"])
for r in records
if r["delta_rses"] is not None
]
correlation = None
if len(paired) >= 3:
xs, ys = zip(*paired)
# Guard: correlation is undefined if either series is constant, which
# happens when a small mock/smoke run produces no variation.
if len(set(xs)) > 1 and len(set(ys)) > 1:
correlation = statistics.correlation(xs, ys)
results = {
"records": records,
"delta_value_vs_delta_rses_correlation": correlation,
"n_paired": len(paired),
}
write_json(out / "claim3.json", results)
write_csv(out / "claim3_records.csv", records)
log(f"claim3: correlation(delta_value, delta_RSES) = {correlation} over n={len(paired)}")
return results
# -----------------------------------------------------------------------------
# Claim 4: pairwise win-rate heatmap (Figure 4).
# -----------------------------------------------------------------------------
def stage_claim4(out: Path, args) -> dict:
"""
Run pairwise comparisons among all methods across the scenario suite.
Args:
out: Output directory.
args: Parsed CLI args.
Returns:
Results dict with the win-rate matrix and every individual verdict.
Why we generate once and compare many times:
Each method produces one transcript per (scenario, seed); all pairwise
comparisons then reuse those transcripts. Regenerating per comparison
would cost O(pairs) episodes instead of O(methods) and would also mean
A-vs-B and A-vs-C judged *different* A transcripts, adding variance that
has nothing to do with the methods.
"""
agent_llm = llm_from_env("agent")
judge = llm_from_env("judge")
rng = random.Random(1234)
scenarios = representative_scenarios()
scenarios += [bullying_scenario("neglectful", n_steps=args.n_steps)]
scenarios += [study_group_scenario(), class_task_scenario(),
election_scenario("neglectful")]
total_available = len(scenarios)
if args.max_scenarios:
scenarios = scenarios[: args.max_scenarios]
# Claim 4 is O(scenarios x methods x seeds) episodes and is by far the most
# expensive stage, so it gets its own seed count. Log any truncation loudly:
# a silently-capped suite would read as "covered everything" when it did not.
n_seeds = args.n_seeds_claim4 or args.n_seeds
if len(scenarios) < total_available:
log(f"claim4: NOTE capped to {len(scenarios)}/{total_available} scenarios "
f"(paper evaluates 17); seeds={n_seeds}")
log(f"claim4: {len(scenarios)} scenarios x {len(METHODS)} methods x {n_seeds} seeds "
f"= {len(scenarios) * len(METHODS) * n_seeds} episodes")
# Stage 1: generate one transcript per (scenario, method, seed).
transcripts: dict[tuple[str, str, int], str] = {}
for scenario in scenarios:
for method in METHODS:
for seed in range(n_seeds):
log(f"claim4: gen {scenario.key}/{method}/{seed}")
try:
ep = run_episode(scenario, method, agent_llm, seed=seed,
concurrency=args.concurrency)
transcripts[(scenario.key, method, seed)] = ep.behaviour_text()
except Exception as exc: # noqa: BLE001
log(f" FAILED: {exc}")
# Stage 2: pairwise judging over the generated transcripts.
verdicts: list[dict] = []
for scenario in scenarios:
for i, m_a in enumerate(METHODS):
for m_b in METHODS[i + 1 :]:
for seed in range(n_seeds):
ta = transcripts.get((scenario.key, m_a, seed))
tb = transcripts.get((scenario.key, m_b, seed))
if not ta or not tb:
continue
winner = pairwise_compare(judge, ta, tb, scenario.setting, rng)
verdicts.append({
"scenario": scenario.key, "method_a": m_a, "method_b": m_b,
"seed": seed,
"winner": {"A": m_a, "B": m_b}.get(winner),
})
# Build the win-rate matrix: matrix[row][col] = win rate of COLUMN vs ROW,
# matching the paper's Figure 4 convention ("the win rate of the column
# model relative to the row model").
matrix: dict[str, dict[str, float]] = {r: {} for r in METHODS}
for row in METHODS:
for col in METHODS:
if row == col:
matrix[row][col] = float("nan")
continue
wins = sum(
1 for v in verdicts
if {v["method_a"], v["method_b"]} == {row, col} and v["winner"] == col
)
losses = sum(
1 for v in verdicts
if {v["method_a"], v["method_b"]} == {row, col} and v["winner"] == row
)
matrix[row][col] = win_rate(wins, losses)
# Average win rate per method across all opponents (the paper's headline
# "strongest overall performance in terms of average win rates").
avg_win = {}
for m in METHODS:
vals = [matrix[r][m] for r in METHODS if r != m and matrix[r][m] == matrix[r][m]]
avg_win[m] = round(statistics.mean(vals), 3) if vals else None
results = {
"matrix": matrix, "average_win_rate": avg_win, "verdicts": verdicts,
"n_scenarios": len(scenarios),
"n_scenarios_available": total_available,
"n_scenarios_paper": 17,
"n_seeds": n_seeds,
"scenarios": [s.key for s in scenarios],
}
write_json(out / "claim4.json", results)
write_csv(out / "claim4_verdicts.csv", verdicts)
log(f"claim4 average win rates: {json.dumps(avg_win)}")
return results
# -----------------------------------------------------------------------------
# Claim 5: election intervention strategies (Figure 7).
# -----------------------------------------------------------------------------
def stage_claim5(out: Path, args) -> dict:
"""
Compare the three intervention strategies against the neglectful control.
Args:
out: Output directory.
args: Parsed CLI args.
Returns:
Results dict with per-arm malicious-competition counts and dispersion.
Why dispersion, not just the mean:
Claim 5 is specifically that interventions "mitigate extreme competitive
tendencies and foster more balanced cooperation", and the paper's
supporting evidence is about spread: "Their lower variance and narrower
ranges across repeated simulations suggest a genuine balancing effect
rather than random fluctuation", with the control showing "the widest
fluctuation". So the testable quantity is the VARIANCE/range of malicious
competition per arm, not only its mean. We record both.
"""
agent_llm = llm_from_env("agent")
judge = llm_from_env("judge")
records: list[dict] = []
for arm in ELECTION_INTERVENTIONS:
scenario = election_scenario(arm, n_agents=args.election_agents,
n_steps=args.n_steps)
for seed in range(args.n_seeds_claim5):
log(f"claim5: arm={arm} seed={seed}")
try:
ep = run_episode(scenario, "EduMirror", agent_llm, seed=seed,
concurrency=args.concurrency)
except Exception as exc: # noqa: BLE001
log(f" FAILED: {exc}")
continue
coded = rate_competition(judge, ep.behaviour_text(), scenario.setting)
# Also validate the Social Value System: does a Surveyor-measured
# SVO angle recover each agent's internal target? (Figure 10.)
svo_rows = []
for persona in scenario.personas:
st = ep.internal_states.get(persona.name, {})
measured = survey_svo(
judge, f"{persona.name}: {persona.description} {persona.goal}",
st.get("svo_target"),
)
svo_rows.append({
"agent": persona.name,
"target": st.get("svo_target"),
"internal_angle": st.get("svo_mean_angle_deg"),
"measured_angle": measured["angle"],
"valid": measured["valid"],
})
records.append({
"arm": arm, "seed": seed,
"cooperative": coded["cooperative"],
"competitive": coded["competitive"],
"malicious_competition": coded["malicious_competition"],
"valid": coded["valid"],
"svo": svo_rows,
})
ep.save(out / "episodes" / f"election_{arm}_{seed}.json")
summary: dict[str, dict] = {}
for arm in ELECTION_INTERVENTIONS:
vals = [r["malicious_competition"] for r in records
if r["arm"] == arm and r["valid"]]
coop = [r["cooperative"] for r in records if r["arm"] == arm and r["valid"]]
if vals:
summary[arm] = {
"n": len(vals),
"malicious_mean": round(statistics.mean(vals), 3),
"malicious_stdev": round(statistics.stdev(vals), 3) if len(vals) > 1 else 0.0,
"malicious_min": min(vals),
"malicious_max": max(vals),
"malicious_range": max(vals) - min(vals),
"cooperative_mean": round(statistics.mean(coop), 3) if coop else None,
}
flat = [{k: v for k, v in r.items() if k != "svo"} for r in records]
results = {"records": records, "summary": summary}
write_json(out / "claim5.json", results)
write_csv(out / "claim5_records.csv", flat)
log(f"claim5 summary: {json.dumps(summary)}")
return results
# -----------------------------------------------------------------------------
# Smoke.
# -----------------------------------------------------------------------------
def stage_smoke(out: Path, args) -> dict:
"""
Tiny end-to-end run over every method and every measurement path.
Args:
out: Output directory.
args: Parsed CLI args.
Returns:
A dict of per-method smoke results.
Why:
Exercises the exact code path the GPU job will take -- every
architecture, the GM loop, the Rater, and the Surveyor -- so plumbing
bugs surface locally in seconds instead of 40 minutes into a paid job.
"""
agent_llm = llm_from_env("agent")
judge = llm_from_env("judge")
scenario = kindergarten_scenario(n_agents=3, n_steps=2)
results = {}
for method in METHODS:
ep = run_episode(scenario, method, agent_llm, seed=0, concurrency=args.concurrency)
rating = rate_behaviour(judge, ep.behaviour_text(), scenario.setting)
results[method] = {
"steps": len(ep.transcript),
"scores": rating.scores,
"average": rating.average(list(RATING_METRICS)),
"usage": ep.usage,
}
log(f"smoke {method}: avg={rating.average(list(RATING_METRICS))}")
write_json(out / "smoke.json", results)
return results
def stage_judgecheck(out: Path, args) -> dict:
"""
Test whether the LLM judge can discriminate quality at all, before trusting it.
Args:
out: Output directory.
args: Parsed CLI args.
Returns:
{"absolute": {...}, "pairwise": {...}} probe results.
Why this runs FIRST:
Our initial GPU run scored every episode exactly 4.0 across all methods
and seeds -- a broken instrument, not a tie. Reporting that as "the
methods are equivalent" would be reporting instrument failure as a
finding. This stage generates one real transcript, corrupts it in ways we
KNOW make it worse (shuffled order, robotic actions), and checks the
judge notices. If it does not, every judged claim downstream is
uninterpretable and the logbook must say so rather than publish numbers.
"""
judge = llm_from_env("judge")
rng = random.Random(7)
scenario = kindergarten_scenario(n_agents=5, n_steps=args.n_steps)
if args.reference_episode:
# JUDGE-ONLY MODE. The probe needs one reference transcript, and it does
# not matter who generated it -- the corruptions are applied to whatever
# it is, and every judge is compared against the SAME text. Reusing an
# archived episode therefore costs zero agent calls, which is what makes
# probing an expensive frontier judge cost cents rather than dollars.
# It also removes a confound: every judge now grades byte-identical text.
ep_data = json.loads(Path(args.reference_episode).read_text())
chunks = []
for entry in ep_data["transcript"]:
chunks.append(f"--- Step {entry['step']} ({entry['location']}) ---")
chunks.append(entry["narration"])
for name, action in entry["actions"].items():
chunks.append(f"{name}: {action}")
transcript = "\n".join(chunks)
log(f"judgecheck: reusing archived reference transcript "
f"({args.reference_episode}, {len(transcript)} chars, "
f"{len(ep_data['transcript'])} steps) -- ZERO agent calls")
else:
agent_llm = llm_from_env("agent")
log("judgecheck: generating one reference transcript (EduMirror) ...")
ep = run_episode(scenario, "EduMirror", agent_llm, seed=0,
concurrency=args.concurrency)
transcript = ep.behaviour_text()
log("judgecheck: probing ABSOLUTE rater (real vs corrupted) ...")
absolute = probe_absolute(judge, transcript, scenario.setting, rng)
log(f" ABSOLUTE verdict={absolute['verdict']} real_avg={absolute['real']['average']} "
f"margins={absolute['margins']}")
log("judgecheck: probing PAIRWISE judge (real vs corrupted) ...")
pairwise = probe_pairwise(judge, transcript, scenario.setting, rng,
n_trials=args.judge_trials)
log(f" PAIRWISE verdict={pairwise['verdict']} "
f"{ {k: v for k, v in pairwise.items() if k in CORRUPTION_KEYS} }")
results = {"absolute": absolute, "pairwise": pairwise,
"reference_transcript_chars": len(transcript),
"judge_usage": judge.usage.as_dict()}
write_json(out / "judgecheck.json", results)
log(f"JUDGECHECK VERDICT: absolute={absolute['verdict']} pairwise={pairwise['verdict']}")
# An UNREADABLE verdict is OUR bug, not a finding about the model. Say so
# loudly and show the evidence, so nobody reads it as "the judge is blind".
for track, res in (("absolute", absolute), ("pairwise", pairwise)):
if res["verdict"] == "UNREADABLE":
log(f" !! {track}: output was UNPARSEABLE -- this is a HARNESS problem, "
f"NOT evidence the judge cannot discriminate.")
for s in res.get("unreadable_samples", [])[:1]:
log(f" raw sample: {s[:400]!r}")
u = judge.usage.as_dict()
if u.get("truncated"):
log(f" !! {u['truncated']}/{u['calls']} judge responses hit max_tokens "
f"(finish_reason=length) -- raise EDUMIRROR_JUDGE_MAX_TOKENS.")
return results
STAGES = {
"smoke": stage_smoke,
"judgecheck": stage_judgecheck,
"claim2": stage_claim2,
"claim3": stage_claim3,
"claim4": stage_claim4,
"claim5": stage_claim5,
}
def main() -> None:
"""Parse args and run the requested stage(s)."""
p = argparse.ArgumentParser(description=__doc__)
p.add_argument("--stage", default="smoke",
choices=list(STAGES) + ["all"], help="which experiment to run")
p.add_argument("--judge-trials", type=int, default=6,
help="pairwise trials per corruption in the judgecheck stage")
p.add_argument("--reference-episode", type=str, default="",
help="path to an archived episode JSON to use as the probe's "
"reference transcript. Enables JUDGE-ONLY mode: no agent "
"calls at all, so probing a paid frontier judge costs "
"cents. Also removes a confound -- every judge then "
"grades byte-identical text.")
p.add_argument("--out", type=Path, default=Path("outputs"))
p.add_argument("--group-sizes", type=int, nargs="+", default=[5, 15, 30],
help="kindergarten group sizes (paper: 5 15 30)")
p.add_argument("--n-steps", type=int, default=8)
p.add_argument("--n-seeds", type=int, default=3,
help="episodes per cell; higher = less noise, more cost")
p.add_argument("--n-seeds-claim5", type=int, default=8,
help="repeats per intervention arm; Claim 5 is about VARIANCE, "
"so it needs more repeats than the mean-based claims")
p.add_argument("--n-seeds-claim4", type=int, default=0,
help="seeds for the pairwise heatmap (0 = use --n-seeds). "
"Claim 4 is the most expensive stage (scenarios x methods "
"x seeds episodes), so it gets its own knob")
p.add_argument("--election-agents", type=int, default=5)
p.add_argument("--max-scenarios", type=int, default=0,
help="cap scenarios in claim4 (0 = all)")
p.add_argument("--concurrency", type=int, default=16)
args = p.parse_args()
out = args.out
out.mkdir(parents=True, exist_ok=True)
stages = ["judgecheck", "claim2", "claim3", "claim4", "claim5"] if args.stage == "all" else [args.stage]
manifest = {"stages": stages, "args": vars(args), "started": time.time()}
t0 = time.time()
for stage in stages:
log(f"=== STAGE {stage} ===")
STAGES[stage](out, args)
manifest["elapsed_s"] = round(time.time() - t0, 1)
write_json(out / "manifest.json", manifest)
log(f"DONE in {manifest['elapsed_s']}s -> {out}")
if __name__ == "__main__":
main()
|