code/self_healing_finetuning/configs/thresholds/model_quality.yaml

paired_eval:
  required: true
  min_samples: 120
  min_candidate_aggregate: 0.60
  min_aggregate_delta_abs: 0.05
  min_candidate_critical_pass_rate: 0.70
  min_critical_pass_delta_abs: 0.00
  max_pairwise_loss_rate: 0.02
  min_pairwise_win_rate: 0.55
  allow_zero_baseline_percent: false
training_budget:
  required: true
  min_train_records: 100
  min_valid_records: 10
  min_max_steps: 300
  require_production_candidate: true
trainer_overfit:
  required: true
  selection_metric: eval_loss
  min_checkpoint_count: 3
  max_train_eval_loss_gap: 0.75
  max_late_eval_loss_regression: 0.10
  require_selected_checkpoint: true
corpus_coverage:
  required: true
  min_total_record_retention: 0.95
  require_train_valid_test: true
  min_train_split_ratio: 0.70
  min_valid_split_ratio: 0.05
  min_test_split_ratio: 0.05
strong_scoring:
  require_model_as_judge: true
  required_rubric_version: model_as_judge_rubric_v1
  min_judged_samples: 30
  min_mean_score: 0.60
  min_critical_pass_rate: 0.70
human_spot_check:
  required: true
  min_reviewed_samples: 10
  require_approval: true
  max_critical_failures: 0
loss_targeted_repair:
  required: false
  min_latest_loss_coverage: 0.90
  min_accepted_repair_targets: 1
  min_preference_pairs: 50