File size: 18,539 Bytes
76bda55 | 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 | """
Bias audit — measures cognitive biases in SGO's LLM evaluator pipeline.
Inspired by CoBRA (Liu et al., CHI'26 Best Paper, arXiv:2509.13588), this script
runs validated social science experiments through SGO's evaluation pipeline to
quantify how much bias the LLM evaluators exhibit.
This is the first step toward expert panel fidelity: you can't calibrate what
you can't measure.
Supported probes:
- framing: same entity, gain vs. loss framing → measures framing effect
- authority: entity with/without authority signals → measures authority bias
- order: same entity, sections reordered → measures anchoring/order effects
Usage:
uv run python scripts/bias_audit.py \
--entity entities/my_product.md \
--cohort data/cohort.json \
--probes framing authority order \
--sample 10 \
--parallel 5
Output: results/bias_audit/report.md + raw data
"""
import json
import os
import re
import time
import argparse
import concurrent.futures
from collections import defaultdict
from datetime import datetime
from pathlib import Path
from dotenv import load_dotenv
PROJECT_ROOT = Path(__file__).resolve().parent.parent
load_dotenv(PROJECT_ROOT / ".env")
from openai import OpenAI
# ── Evaluation core (reused from evaluate.py) ────────────────────────────
SYSTEM_PROMPT = """You are an evaluation simulator. You will be given:
1. A detailed persona — a person with specific values, needs, context, and perspective
2. An entity to evaluate (a product, profile, proposal, pitch, resume, etc.)
Your job: fully inhabit this persona's perspective and evaluate the entity AS THEY WOULD.
Be honest and realistic. Not everything is a match. Consider:
- Their specific needs, budget, constraints, and priorities
- Whether this entity solves a real problem for them
- Trust signals and red flags from their perspective
- Practical fit with their situation
- What they'd compare this against
You MUST respond with valid JSON only."""
EVAL_PROMPT = """## Evaluator Persona
Name: {name}
Age: {age}
Location: {city}, {state}
Education: {education_level}
Occupation: {occupation}
Status: {marital_status}
{persona}
---
## Entity to Evaluate
{entity}
---
## Task
Inhabit {name}'s perspective completely. Evaluate this entity as they would.
Return JSON:
{{
"score": <1-10, where 1=strong reject, 5=ambivalent, 10=enthusiastic yes>,
"action": "<positive | neutral | negative>",
"attractions": ["<what works for them, max 3>"],
"concerns": ["<what gives them pause, max 3>"],
"dealbreakers": ["<hard no's if any, empty list if none>"],
"summary": "<1-2 sentences — how they'd describe this to a peer>",
"reasoning": "<2-3 sentence internal monologue>"
}}"""
def evaluate_one(client, model, evaluator, entity_text):
prompt = EVAL_PROMPT.format(
name=evaluator["name"],
age=evaluator.get("age", ""),
city=evaluator.get("city", ""),
state=evaluator.get("state", ""),
education_level=evaluator.get("education_level", ""),
occupation=evaluator.get("occupation", ""),
marital_status=evaluator.get("marital_status", ""),
persona=evaluator.get("persona", ""),
entity=entity_text,
)
try:
resp = client.chat.completions.create(
model=model,
messages=[
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": prompt},
],
response_format={"type": "json_object"},
max_tokens=16384,
temperature=0.7,
)
content = resp.choices[0].message.content
if not content:
return {"error": "Empty response"}
content = re.sub(r'<think>[\s\S]*?</think>', '', content).strip()
result = json.loads(content)
result["_evaluator"] = {
"name": evaluator["name"],
"age": evaluator.get("age"),
"occupation": evaluator.get("occupation"),
}
return result
except Exception as e:
return {"error": str(e), "_evaluator": {"name": evaluator.get("name", "?")}}
# ── Bias probes ──────────────────────────────────────────────────────────
REFRAME_PROMPT = """You are a text transformation tool. Rewrite the following entity description
using {frame_type} framing. Keep ALL factual content identical — same features, same pricing,
same capabilities. Only change the rhetorical framing.
{frame_instruction}
Return the rewritten text only, no commentary.
---
{entity}"""
FRAME_INSTRUCTIONS = {
"gain": "Emphasize what the user GAINS: benefits, improvements, positive outcomes. "
'Use phrases like "save", "gain", "achieve", "unlock", "improve".',
"loss": "Emphasize what the user LOSES without this: risks, costs of inaction, missed opportunities. "
'Use phrases like "avoid losing", "stop wasting", "don\'t miss", "risk of", "falling behind".',
}
def reframe_entity(client, model, entity_text, frame_type):
"""Rewrite entity with gain or loss framing, preserving factual content."""
resp = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": REFRAME_PROMPT.format(
frame_type=frame_type,
frame_instruction=FRAME_INSTRUCTIONS[frame_type],
entity=entity_text,
)}],
max_tokens=16384,
temperature=0.3,
)
return resp.choices[0].message.content.strip()
AUTHORITY_SIGNALS = [
"Trusted by 10,000+ teams worldwide.",
"SOC 2 Type II certified. GDPR compliant.",
"Featured in TechCrunch, Wired, and Forbes.",
"Backed by Sequoia Capital and Y Combinator.",
"Winner of Product Hunt's Product of the Year.",
]
def add_authority_signals(entity_text):
"""Add authority/credibility signals to an entity."""
signals = "\n".join(f"- {s}" for s in AUTHORITY_SIGNALS)
return f"{entity_text}\n\n---\n\n### Trust & Recognition\n\n{signals}\n"
def reorder_entity(entity_text):
"""Reverse the order of sections in the entity document."""
sections = re.split(r'\n(?=##?\s)', entity_text)
if len(sections) <= 1:
# Try splitting on blank lines if no headers
sections = re.split(r'\n\n+', entity_text)
if len(sections) <= 1:
return entity_text # Can't reorder a single section
# Keep first section (title/intro), reverse the rest
return sections[0] + "\n\n" + "\n\n".join(reversed(sections[1:]))
# ── Probe runners ────────────────────────────────────────────────────────
def run_paired_evaluation(client, model, evaluators, entity_a, entity_b, label_a, label_b, parallel):
"""Run the same cohort against two entity variants and compute deltas."""
results = []
def worker(ev):
r_a = evaluate_one(client, model, ev, entity_a)
r_b = evaluate_one(client, model, ev, entity_b)
return {
"evaluator": ev["name"],
"age": ev.get("age"),
"occupation": ev.get("occupation"),
f"score_{label_a}": r_a.get("score"),
f"score_{label_b}": r_b.get("score"),
"delta": (r_b.get("score", 0) or 0) - (r_a.get("score", 0) or 0),
f"reasoning_{label_a}": r_a.get("reasoning", ""),
f"reasoning_{label_b}": r_b.get("reasoning", ""),
"error": r_a.get("error") or r_b.get("error"),
}
done = [0]
with concurrent.futures.ThreadPoolExecutor(max_workers=parallel) as pool:
futs = {pool.submit(worker, ev): ev for ev in evaluators}
for fut in concurrent.futures.as_completed(futs):
result = fut.result()
results.append(result)
done[0] += 1
if result.get("error"):
print(f" [{done[0]}/{len(evaluators)}] {result['evaluator']}: ERROR")
else:
print(f" [{done[0]}/{len(evaluators)}] {result['evaluator']}: "
f"{label_a}={result[f'score_{label_a}']} "
f"{label_b}={result[f'score_{label_b}']} "
f"Δ={result['delta']:+d}")
return results
def run_framing_probe(client, model, evaluators, entity_text, parallel):
"""Framing Effect probe: gain-framed vs. loss-framed entity."""
print("\n── Framing Effect Probe ──")
print("Generating gain-framed and loss-framed variants...")
gain_entity = reframe_entity(client, model, entity_text, "gain")
loss_entity = reframe_entity(client, model, entity_text, "loss")
return run_paired_evaluation(
client, model, evaluators, gain_entity, loss_entity,
"gain", "loss", parallel,
), {"gain_entity": gain_entity, "loss_entity": loss_entity}
def run_authority_probe(client, model, evaluators, entity_text, parallel):
"""Authority Bias probe: entity with vs. without authority signals."""
print("\n── Authority Bias Probe ──")
entity_with_authority = add_authority_signals(entity_text)
return run_paired_evaluation(
client, model, evaluators, entity_text, entity_with_authority,
"baseline", "authority", parallel,
), {"entity_with_authority": entity_with_authority}
def run_order_probe(client, model, evaluators, entity_text, parallel):
"""Order Effect probe: original vs. reordered entity."""
print("\n── Order Effect Probe ──")
reordered = reorder_entity(entity_text)
return run_paired_evaluation(
client, model, evaluators, entity_text, reordered,
"original", "reordered", parallel,
), {"reordered_entity": reordered}
# ── Analysis ─────────────────────────────────────────────────────────────
HUMAN_BASELINES = {
"framing": {
"description": "Tversky & Kahneman (1981): ~30% of subjects shift preference based on framing",
"expected_shift_pct": 30,
},
"authority": {
"description": "Milgram (1963): 65% obedience rate under authority pressure",
"expected_shift_pct": 20, # Conservative estimate for evaluation context
},
"order": {
"description": "Primacy/recency effects: ideally 0% shift (order shouldn't matter)",
"expected_shift_pct": 0,
},
}
def analyze_probe(results, probe_name, label_a, label_b):
"""Analyze a probe's results and compare to human baselines."""
valid = [r for r in results if not r.get("error")]
if not valid:
return {"probe": probe_name, "error": "No valid results"}
deltas = [r["delta"] for r in valid]
abs_deltas = [abs(d) for d in deltas]
shifted = [r for r in valid if r["delta"] != 0]
positive_shift = [r for r in valid if r["delta"] > 0]
negative_shift = [r for r in valid if r["delta"] < 0]
n = len(valid)
avg_delta = sum(deltas) / n
avg_abs_delta = sum(abs_deltas) / n
shift_pct = 100 * len(shifted) / n
baseline = HUMAN_BASELINES.get(probe_name, {})
return {
"probe": probe_name,
"n": n,
"avg_delta": round(avg_delta, 2),
"avg_abs_delta": round(avg_abs_delta, 2),
"max_delta": max(deltas),
"min_delta": min(deltas),
"shifted_pct": round(shift_pct, 1),
"positive_shifts": len(positive_shift),
"negative_shifts": len(negative_shift),
"no_change": n - len(shifted),
"human_baseline": baseline,
"comparison": label_a + " vs " + label_b,
}
def generate_report(all_analyses, model):
"""Generate the bias audit report."""
lines = [
"# SGO Bias Audit Report",
f"\n**Date**: {datetime.now().isoformat()}",
f"**Model**: {model}",
f"**Method**: CoBRA-inspired social science experiments (arXiv:2509.13588)",
"",
"---",
"",
"## Summary",
"",
f"{'Probe':<12} {'N':>4} {'Avg Δ':>7} {'|Δ|':>5} {'Shifted%':>9} {'Human Baseline':>15} Gap",
"-" * 75,
]
for a in all_analyses:
if "error" in a:
lines.append(f"{a['probe']:<12} ERROR: {a['error']}")
continue
baseline_pct = a["human_baseline"].get("expected_shift_pct", "?")
gap = ""
if isinstance(baseline_pct, (int, float)):
diff = a["shifted_pct"] - baseline_pct
gap = f"{diff:+.1f}pp"
lines.append(
f"{a['probe']:<12} {a['n']:>4} {a['avg_delta']:>+6.2f} {a['avg_abs_delta']:>5.2f}"
f" {a['shifted_pct']:>5.1f}% {str(baseline_pct)+('%' if isinstance(baseline_pct, (int,float)) else ''):>15} {gap}"
)
lines.extend(["", "---", "", "## Interpretation", ""])
for a in all_analyses:
if "error" in a:
continue
lines.append(f"### {a['probe'].title()} Effect ({a['comparison']})")
lines.append("")
baseline = a["human_baseline"]
if baseline:
lines.append(f"**Human baseline**: {baseline.get('description', 'N/A')}")
lines.append(f"**LLM result**: {a['shifted_pct']:.1f}% of evaluators shifted scores "
f"(avg |Δ| = {a['avg_abs_delta']:.2f} points)")
expected = baseline.get("expected_shift_pct")
if isinstance(expected, (int, float)):
if a["shifted_pct"] > expected + 10:
lines.append(f"**Assessment**: OVER-BIASED — LLM evaluators show more {a['probe']} "
f"sensitivity than humans. Consider adding de-biasing instructions.")
elif a["shifted_pct"] < expected - 10:
lines.append(f"**Assessment**: UNDER-BIASED — LLM evaluators show less {a['probe']} "
f"sensitivity than humans. The panel may be too rational.")
else:
lines.append(f"**Assessment**: WELL-CALIBRATED — within ±10pp of human baseline.")
lines.append("")
lines.extend([
"---",
"",
"## Next Steps",
"",
"1. **If over-biased**: Add bias-awareness instructions to the evaluation prompt",
"2. **If under-biased**: Consider if this is acceptable (more rational) or needs calibration",
"3. **For order effects**: Any non-zero shift indicates entity structure matters — "
"standardize entity format or average across orderings",
"4. **Re-run after calibration**: Use this script to verify improvements",
"",
"## References",
"",
"- Liu, X., Shang, H., & Jin, H. (2025). CoBRA. arXiv:2509.13588 (CHI'26 Best Paper)",
"- Tversky, A. & Kahneman, D. (1981). The framing of decisions. Science, 211(4481).",
"- Milgram, S. (1963). Behavioral Study of Obedience. JASP, 67(4).",
])
return "\n".join(lines)
# ── Main ─────────────────────────────────────────────────────────────────
def main():
parser = argparse.ArgumentParser(description="Bias audit for SGO evaluator pipeline")
parser.add_argument("--entity", required=True, help="Path to entity document")
parser.add_argument("--cohort", default="data/cohort.json")
parser.add_argument("--probes", nargs="+", default=["framing", "authority", "order"],
choices=["framing", "authority", "order"])
parser.add_argument("--sample", type=int, default=10,
help="Number of evaluators to sample for audit (smaller = faster)")
parser.add_argument("--parallel", type=int, default=5)
args = parser.parse_args()
entity_text = Path(args.entity).read_text()
client = OpenAI(api_key=os.getenv("LLM_API_KEY"), base_url=os.getenv("LLM_BASE_URL"))
model = os.getenv("LLM_MODEL_NAME")
with open(args.cohort) as f:
cohort = json.load(f)
# Sample a subset for the audit (bias audit is 2x cost per evaluator per probe)
import random
random.seed(42)
if args.sample and args.sample < len(cohort):
evaluators = random.sample(cohort, args.sample)
else:
evaluators = cohort
print(f"Bias Audit | {len(evaluators)} evaluators | Model: {model}")
print(f"Probes: {', '.join(args.probes)}")
probe_runners = {
"framing": lambda: run_framing_probe(client, model, evaluators, entity_text, args.parallel),
"authority": lambda: run_authority_probe(client, model, evaluators, entity_text, args.parallel),
"order": lambda: run_order_probe(client, model, evaluators, entity_text, args.parallel),
}
all_results = {}
all_analyses = []
for probe_name in args.probes:
t0 = time.time()
results, metadata = probe_runners[probe_name]()
elapsed = time.time() - t0
label_a, label_b = {
"framing": ("gain", "loss"),
"authority": ("baseline", "authority"),
"order": ("original", "reordered"),
}[probe_name]
analysis = analyze_probe(results, probe_name, label_a, label_b)
analysis["elapsed_s"] = round(elapsed, 1)
all_analyses.append(analysis)
all_results[probe_name] = {
"results": results,
"metadata": metadata,
"analysis": analysis,
}
print(f"\n {probe_name}: avg Δ={analysis.get('avg_delta', '?'):+.2f}, "
f"shifted={analysis.get('shifted_pct', '?')}%, "
f"time={elapsed:.1f}s")
# Save
out_dir = PROJECT_ROOT / "results" / "bias_audit"
out_dir.mkdir(parents=True, exist_ok=True)
# Raw data
serializable = {}
for k, v in all_results.items():
serializable[k] = {
"results": v["results"],
"analysis": v["analysis"],
}
with open(out_dir / "raw_data.json", "w") as f:
json.dump(serializable, f, ensure_ascii=False, indent=2)
# Report
report = generate_report(all_analyses, model)
with open(out_dir / "report.md", "w") as f:
f.write(report)
print(f"\nReport: {out_dir / 'report.md'}")
print(f"Data: {out_dir / 'raw_data.json'}")
print(f"\n{report}")
if __name__ == "__main__":
main()
|