{
  "schema": "albedo-sgp-dpo-summary-v1",
  "counts": {
    "pairs:dev": 1237,
    "pairs:label_source:legacy_teacher_default": 5972,
    "pairs:label_source:measured_exact_paired": 387,
    "pairs:label_source:measured_onepass": 5893,
    "pairs:preference:king": 1558,
    "pairs:preference:teacher": 10081,
    "pairs:preference:tie": 613,
    "pairs:train": 11015,
    "samples:dev": 7414,
    "samples:k1": 12252,
    "samples:k2": 12252,
    "samples:k3": 12252,
    "samples:k4": 12252,
    "samples:k5": 12239,
    "samples:k6": 12157,
    "samples:train": 65990
  },
  "pairs": 12252,
  "samples": 73404,
  "samples_per_pair_mean": 5.991,
  "tie_share_of_pairs": 0.05,
  "selection": {
    "leading_groups_that_must_match": 1,
    "later_groups_unconstrained": true,
    "legacy_label": "teacher is chosen by default (pre-prune self-score, not a measurement)",
    "measured_label": "follows the judged score on the same frozen checklist",
    "tie_target_share": 0.05,
    "legacy_pathologies_dropped": 3188,
    "legacy_pathology_rule": "env_burn (4+ consecutive environment turns) or no_verification"
  },
  "turn_split": {
    "format": "progressive prefixes; sample k compares the first k behaviour groups",
    "prompt_is_shared": true,
    "label_measured_on": "complete trajectory",
    "filter_to_one_sample_per_pair": "is_full_trajectory == true"
  },
  "loss_contract": {
    "assistant_tokens_of_the_final_group_only": true,
    "earlier_groups_are_unsupervised_context": true,
    "observations_are_context": true,
    "per_message_loss_flag": "loss",
    "each_group_is_a_target_exactly_once_across_k": true
  }
}