{
  "schema_version": "1.0",
  "experiment_id": "multihop-rag-orthogonal-residual-subtraction-benchmarks",
  "origin": "first_party_lab",
  "recorded_at": "2026-10-03",
  "environment": {
    "corpus": "HotpotQA Wikipedia (66,635 documents)",
    "evaluated_queries_n": 1395,
    "step1_top1_hit_subset_n": 1029,
    "compute_runtime": "Python 3.11 / PyTorch (CPU 16-thread batch evaluation + float32 causal state verification)",
    "trajectory_pilot_model": "Qwen2.5-0.5B base (448 disjoint documents, 36,435 intermediate token states, 130.58 MB)"
  },
  "probe_reach_rate_metrics_n1029": {
    "global_orthogonal_residual_r": {
      "bridge_reach_at_10_pct": 18.45,
      "bridge_reach_at_50_pct": 27.31,
      "comparison_reach_at_10_pct": 83.26,
      "comparison_reach_at_50_pct": 92.07,
      "overall_reach_at_10_pct": 32.75,
      "overall_reach_at_50_pct": 41.59
    },
    "first_hop_document_d1_only": {
      "bridge_reach_at_10_pct": 41.9,
      "bridge_reach_at_50_pct": 57.23,
      "comparison_reach_at_10_pct": 4.85,
      "comparison_reach_at_50_pct": 12.33,
      "overall_reach_at_10_pct": 33.72,
      "overall_reach_at_50_pct": 47.33
    },
    "local_sentence_residual_only": {
      "bridge_reach_at_10_pct": 24.31,
      "bridge_reach_at_50_pct": 34.54,
      "comparison_reach_at_50_pct": 92.07,
      "overall_reach_at_50_pct": 47.23
    },
    "global_blend_r_plus_05_d1": {
      "bridge_reach_at_50_pct": 49.88,
      "comparison_reach_at_50_pct": 94.27,
      "overall_reach_at_50_pct": 59.67
    },
    "global_blend_r_plus_10_d1": {
      "bridge_reach_at_10_pct": 46.26,
      "bridge_reach_at_50_pct": 61.22,
      "comparison_reach_at_50_pct": 89.43,
      "overall_reach_at_50_pct": 67.44
    },
    "local_residual_plus_10_d1_hybrid": {
      "bridge_reach_at_10_pct": 50.0,
      "bridge_reach_at_50_pct": 62.84,
      "overall_reach_at_50_pct": 68.32,
      "gain_over_global_blend_at_50_pt": 1.62
    },
    "quad_probe_union_budget_36": {
      "bridge_reach_at_50_pct": 66.08,
      "comparison_reach_at_50_pct": 96.04,
      "overall_reach_at_50_pct": 72.69
    }
  },
  "negative_results_and_audit_diagnostics": {
    "multi_slot_representation_collapse_exp31": {
      "sampled_documents_n": 100,
      "inter_slot_cosine_similarity_mean": 0.99900543,
      "inter_slot_cosine_similarity_min": 0.99034917,
      "max_slot_component_diff": 0.0277777,
      "max_slot_attention_diff": 0.16343284,
      "unconnected_query_projection_params": "4 / 4",
      "train_test_gold_doc_overlap_found": 293
    },
    "autoregressive_trajectory_hidden_state_pilot_h1": {
      "disjoint_documents_n": 448,
      "evaluated_queries_n": 32,
      "bootstrap_resamples": 10000,
      "intermediate_states_captured": 36435,
      "mrr_diff_vs_document_text": -0.84780663,
      "mrr_diff_95ci_vs_document_text": [
        -0.91435415,
        -0.76957864
      ],
      "any_recall5_diff_vs_document_text": -0.8125,
      "full_match2_diff_vs_document_text": -0.625,
      "mrr_diff_vs_final_hidden_state": 0.04265682,
      "hypothesis_h1_supported": false
    }
  }
}
