Calibration: run web-dev-html-001 eval against Claude and record first scores

[Lingz450]

Objective

Run the calibration script against all 5 eval samples and compare AI scores to expected ranges.

Steps

1. Add ANTHROPIC_API_KEY to backend/.env
2. Start the server: uvicorn app.main:app --reload --port 8000
3. Run: python evals/run_calibration.py --eval evals/web-dev-html-001.jsonl
4. Record which samples PASS and which FAIL

Definition of done

• All 5 samples return a score within their expected_score_range
• Results saved to evals/calibration_results/
• If any sample fails, rubric scoring_guide is adjusted and re-run until pass rate >= 80%

Files involved

• evals/web-dev-html-001.jsonl — the 5 calibration samples
• evals/run_calibration.py — the runner
• rubrics/web-dev-html-001.json — the rubric to tune if needed