Add CJE integration for calibrated evaluation

[elandesberg]

Summary

Adds integration with CJE (Causal Judge Evaluation) to enable calibrated evaluation of AI survey responses against human ground truth.

Key benefits:

• Calibrate with 5-10% human labels to get valid estimates for the full dataset
• Proper uncertainty quantification with confidence intervals
• Statistical comparison between models/policies

Usage

# Run survey with multiple models
results = survey.by([Model("gpt-4o"), Model("claude-3-5-sonnet")]).run()

# Add human labels for ~10% of responses
results = results.add_column("human_rating", human_labels)

# Calibrate
cal = results.calibrate("sentiment_score", "human_rating")

print(cal.estimates)
# {'gpt-4o': 0.72, 'claude-3-5-sonnet': 0.68}

print(cal.confidence_intervals)
# {'gpt-4o': (0.68, 0.76), 'claude-3-5-sonnet': (0.64, 0.72)}

# Statistical comparison
cal.compare("gpt-4o", "claude-3-5-sonnet")
# ComparisonResult(diff=0.04, p=0.046, significant=True)

Install: pip install edsl[cje]

Changes

• New edsl/cje_integration/ package (calibrator, data adapters, result types)
• Added Results.calibrate() convenience method
• Added cje-eval as optional dependency with [cje] extra
• Demo notebook: docs/notebooks/cje_calibration_demo.ipynb
• 11 passing tests

Test plan

• All 11 unit tests pass (pytest tests/cje_integration/ -v)
• End-to-end test with real CJE passes
• Review demo notebook runs correctly

[elandesberg]

Any feedback is appreciated! Just a first pass - hope it makes sense.

[elandesberg]

Actually, I just ran into a bug in the notebook, so please hold off on review.

[elandesberg]

Ok, this is ready for a review. No rush. Thank you.