{
  "source_sha256": "6b1d0206fab586110b248b85ff89e2e9d88dd834745555c344fade0e38d203a1",
  "planned_cells": 480,
  "valid": 425,
  "excluded": 55,
  "isolated_regraded": 425,
  "regrade_errors": {},
  "grade_changes": [],
  "judge_counts": {
    "codex:valid": 168,
    "claude:error": 48,
    "claude:valid": 120
  },
  "models_observed": [
    "claude-sonnet-5-5"
  ],
  "total_ledger_reported_usd": 16.1344918,
  "method": "Exploratory, equal task weight; 2000 within-task bootstrap draws of BOTH baseline and treatment. Fixed task suite; no claim of generalization across task populations. Existing main_effects.csv is superseded for public inference."
}