{
  "format": "chromatic-seed-held-out/1",
  "completed_utc": "2026-10-08T12:26:00.003794+00:00",
  "first_test_read_utc": "2026-10-08T12:25:52.270234+00:00",
  "checkpoint_pin": {
    "pinned_utc_before_test_read": "2026-10-08T12:25:52.262303+00:00",
    "selected_update": 200,
    "selected_weights_sha256": "9b4e05d498fee7c16b0145bed04fb6dad52e76a86857c08c101f467d9f62d118",
    "initial_weights_sha256": "ed4371f1bb563880ef6940335809defc497abe32b730bfa17f114d810b18df5c",
    "files": {
      "runs/seed01/best.npz": "8a34b2f60d0ef5a89d22b2afc3c421547979cd3471d0c2d1371bc963bb426969",
      "runs/seed01/initial.npz": "f8146cbe9502cfe672de3ed191f9ba7ff8d85f1e18f0d3ca806c58ea5ace2625",
      "runs/seed01/summary.json": "a29aafa8b63116305f3ca0c4723dedd30e6d469ea755bc6410eb75e1c5a8f71a",
      "protocol/config.json": "14fcabb37bfdba73e3d3411d07ab2962b365bf75f4fb3c3aa708e8efc5fbe183",
      "protocol/PROTOCOL.txt": "bfefdba4e108e4538973770798dd9fe83be405c3051756253b3331236d80abae",
      "tools/evaluate_seed.py": "4fbf38df5715181ab2a6872feb44ba59b28e2b8be94c092493a79a1035629d66"
    },
    "selection_fixed_before_test_read": true,
    "methods": [
      "uniform",
      "training_frequency",
      "training_one_hop",
      "training_two_hop",
      "initial_model",
      "selected_model"
    ],
    "vocabulary_size": 6257,
    "two_hop_fallback": "Missing second-hop branch mass goes uniformly to endpoints; dead source is uniform",
    "tie_policy": "Exact equality; uniformly random order within the tied group"
  },
  "queries": 2038,
  "scored_queries": 2038,
  "excluded_queries": 0,
  "vocabulary_size": 6257,
  "candidate_restriction": null,
  "selected_update": 200,
  "training_scored_tokens": 318308,
  "target_observed_in_training_queries": 1839,
  "target_unseen_in_training_queries": 199,
  "test_source_without_training_outgoing_queries": 115,
  "metrics": {
    "uniform": {
      "all": {
        "queries": 2038,
        "nll": 8.741456115998362,
        "top1": 0.000159821000479463,
        "top10": 0.0015982100047946292,
        "mrr": 0.0014893322181990627,
        "expected_rank": 3129.0,
        "target_probability": 0.000159821000479463
      },
      "target_observed_in_training": {
        "queries": 1839,
        "nll": 8.741456115998362,
        "top1": 0.00015982100047946297,
        "top10": 0.0015982100047946294,
        "mrr": 0.0014893322181990627,
        "expected_rank": 3129.0,
        "target_probability": 0.00015982100047946297
      },
      "target_unseen_in_training": {
        "queries": 199,
        "nll": 8.741456115998364,
        "top1": 0.00015982100047946303,
        "top10": 0.0015982100047946294,
        "mrr": 0.0014893322181990622,
        "expected_rank": 3129.0,
        "target_probability": 0.00015982100047946303
      }
    },
    "training_frequency": {
      "all": {
        "queries": 2038,
        "nll": 8.97132209244305,
        "top1": 0.018645731108930325,
        "top10": 0.0323846908734053,
        "mrr": 0.02645789819733658,
        "expected_rank": 1410.4401373895976,
        "target_probability": 0.0017813122771541715
      },
      "target_observed_in_training": {
        "queries": 1839,
        "nll": 8.568894206011654,
        "top1": 0.02066340402392605,
        "top10": 0.03588907014681892,
        "mrr": 0.02929917085694175,
        "expected_rank": 1014.7601957585645,
        "target_probability": 0.0019737364287651937
      },
      "target_unseen_in_training": {
        "queries": 199,
        "nll": 12.690241103233692,
        "top1": 0.0,
        "top10": 0.0,
        "mrr": 0.00020111216209083322,
        "expected_rank": 5067.0,
        "target_probability": 3.081046939750128e-06
      }
    },
    "training_one_hop": {
      "all": {
        "queries": 2038,
        "nll": 10.914111235638966,
        "top1": 9.018358712040356e-06,
        "top10": 0.0012229866335313948,
        "mrr": 0.0012494481483324913,
        "expected_rank": 3130.175171736997,
        "target_probability": 2.4098622888782622e-05
      },
      "target_observed_in_training": {
        "queries": 1839,
        "nll": 10.906311812456604,
        "top1": 9.559711828570381e-06,
        "top10": 0.001228742540281831,
        "mrr": 0.001251855308879102,
        "expected_rank": 3130.15715062534,
        "target_probability": 2.4585840693659645e-05
      },
      "target_unseen_in_training": {
        "queries": 199,
        "nll": 10.986187312183516,
        "top1": 4.015603027122186e-06,
        "top10": 0.0011697951133602794,
        "mrr": 0.0012272030817736118,
        "expected_rank": 3130.3417085427136,
        "target_probability": 1.959614277235627e-05
      }
    },
    "training_two_hop": {
      "all": {
        "queries": 2038,
        "nll": 8.91203747296301,
        "top1": 0.056014331444148385,
        "top10": 0.206594069938707,
        "mrr": 0.10073271032860921,
        "expected_rank": 2467.2075564278703,
        "target_probability": 0.037404164553794056
      },
      "target_observed_in_training": {
        "queries": 1839,
        "nll": 8.702055926099444,
        "top1": 0.06207517594190948,
        "top10": 0.22888425221111938,
        "mrr": 0.11151282574291574,
        "expected_rank": 2395.2191408374115,
        "target_probability": 0.04144898844912849
      },
      "target_unseen_in_training": {
        "queries": 199,
        "nll": 10.852520210058996,
        "top1": 4.818723632546623e-06,
        "top10": 0.0006059031097303209,
        "mrr": 0.0011114427561988009,
        "expected_rank": 3132.467336683417,
        "target_probability": 2.5113581331622158e-05
      }
    },
    "initial_model": {
      "all": {
        "queries": 2038,
        "nll": 8.747568545180298,
        "top1": 0.0004906771344455348,
        "top10": 0.0014720314033366045,
        "mrr": 0.001744650426682714,
        "expected_rank": 3087.8846908734054,
        "target_probability": 0.0001604044157003707
      },
      "target_observed_in_training": {
        "queries": 1839,
        "nll": 8.749005131573156,
        "top1": 0.000543773790103317,
        "top10": 0.0016313213703099511,
        "mrr": 0.0017902698053149212,
        "expected_rank": 3105.7911908646,
        "target_probability": 0.0001601662824585456
      },
      "target_unseen_in_training": {
        "queries": 199,
        "nll": 8.734292754343796,
        "top1": 0.0,
        "top10": 0.0,
        "mrr": 0.0013230723497750308,
        "expected_rank": 2922.407035175879,
        "target_probability": 0.00016260505405070425
      }
    },
    "selected_model": {
      "all": {
        "queries": 2038,
        "nll": 8.655722640783505,
        "top1": 0.005397448478900883,
        "top10": 0.02404317958783121,
        "mrr": 0.014574231540494777,
        "expected_rank": 1540.031894013739,
        "target_probability": 0.0014441069639837868
      },
      "target_observed_in_training": {
        "queries": 1839,
        "nll": 8.429496691339677,
        "top1": 0.005981511691136488,
        "top10": 0.026644915715062535,
        "mrr": 0.01612604113027115,
        "expected_rank": 1215.0690592713431,
        "target_probability": 0.0015959437465454507
      },
      "target_unseen_in_training": {
        "queries": 199,
        "nll": 10.746323248960381,
        "top1": 0.0,
        "top10": 0.0,
        "mrr": 0.00023363940180758794,
        "expected_rank": 4543.08040201005,
        "target_probability": 4.0951973376247884e-05
      }
    }
  },
  "best_nll_method": "selected_model",
  "best_mrr_method": "training_two_hop",
  "selected_minus_initial_nll": -0.09184590439679319,
  "selected_minus_uniform_nll": -0.08573347521485708,
  "selected_minus_frequency_nll": -0.3155994516595442,
  "probability_floor_for_baselines": 2.2250738585072014e-308,
  "model_nll": "Float64 log-sum-exp of saved float32 full-vocabulary logits; no truncation",
  "frequency_baseline": "All non-ignored training target occurrences, add-one over full vocabulary",
  "graph_baselines": "Uniform outgoing training-neighbor draws; 0.9 graph plus 0.1 uniform; rational two-hop mass accumulation",
  "tie_metrics": "Expected top-k and MRR under uniform ordering within exact score ties; inclusive one-based rank bounds",
  "stratification": "Descriptive target-observed vs target-unseen training strata; no queries filtered",
  "inference_device": "cuda",
  "torch": "2.7.1+cu126",
  "numpy": "2.3.5",
  "gpu_max_memory_bytes": 63720960,
  "elapsed_seconds": 7.7729107000050135,
  "method_seconds": {
    "uniform": 0.07302740000886843,
    "training_frequency": 0.07765019999351352,
    "training_one_hop": 0.1282226000330411,
    "training_two_hop": 0.15736289997585118,
    "initial_model": 0.5300690000294708,
    "selected_model": 0.25087129999883473
  },
  "input_sha256": {
    "data/test.npz": "08eeb65df5aff1eac5ed837f826a8ec8ebfacf0b6ead5cfd260a444f703fc167",
    "data/train.npz": "1b8619ac00fb67610c7c5110c704bf13d54389f74304403ac66e315be7e130f8",
    "data/edges.npz": "ee82465e71b42b17341d9fc9a2ce8662288d203cbf4ddae043938ee337233248"
  },
  "per_query_archive": {
    "path": "analysis/evaluation-queries.npz",
    "bytes": 102554813,
    "sha256": "d7a58db57976a421e446e574d50902b0762cbe3f22108322a281c05da23aee64",
    "contents": "Each query target/context, all per-method target probabilities/NLL/rank bounds/expected rank/top1/top10/MRR, complete model logits and complete baseline probabilities"
  },
  "limitations": [
    "Single fixed graph snapshot and one initialization; dependent graph queries.",
    "No test tuning, exclusion, population significance test, or semantic-understanding inference.",
    "Target-observed strata describe exposure and are not independent experimental groups."
  ]
}