next_experiment_recommendation.json

{
  "milestone": "L3-43 v0.1 evidence milestone",
  "data_version": "l3_20260507_43case",
  "status": "ready_for_result_discussion",
  "do_not_do_now": [
    "do not expand data",
    "do not restart completed L3-43 guard pipeline",
    "do not claim final foundation-model status"
  ],
  "recommended_order": [
    {
      "rank": 1,
      "experiment": "structure_context_m6",
      "decision": "run_next",
      "rationale": "Full M6 underperforms baseline on label retrieval and the current L3-43 objective did not optimize structure context.",
      "success_criteria": {
        "label_top1_or_top5": "above current Full M6",
        "gene_correlation": ">=0.995",
        "pointer_errors": 0
      }
    },
    {
      "rank": 2,
      "experiment": "virchow_uni_multi_case_smoke",
      "decision": "run_after_failure_review",
      "rationale": "Only useful if failure review suggests lightweight image encoder capacity is limiting structure or morphology semantics.",
      "success_criteria": {
        "image_gene_top1": ">=0.90 on smoke",
        "label_retrieval": "improves over lightweight image encoder smoke"
      }
    },
    {
      "rank": 3,
      "experiment": "data_expansion",
      "decision": "defer",
      "rationale": "The existing 43-case milestone is sufficient for objective and encoder decisions."
    }
  ],
  "diagnostics": {
    "full_label_top1": 0.06296011277657843,
    "baseline_label_top1": 0.12827654778362696,
    "baseline_minus_full_label_top1": 0.06531643500704853,
    "pointer_errors_total": 0,
    "full_failure_rows": 18,
    "baseline_failure_rows": 18
  }
}