{
  "version": 2,
  "question": "How far does each single wrong label reach, compared with k-NN(3) and logistic regression on the identical inputs?",
  "release": {
    "name": "NVIDIA Kumo Tabular",
    "date": "2026-09-29",
    "url": "https://huggingface.co/blog/nvidia/kumo-tabular"
  },
  "source_revision": "391961ae82b4bf54af1aca31e9fc0a0c678117ce",
  "model_revision": "3c3e10bbdb590ace29e7026847f92db3c603096d",
  "model_file": "small/classifier.pt",
  "model_license": "OpenMDW-1.1",
  "seed": 20261006,
  "context_axis": [
    -0.9,
    -0.3,
    0.3,
    0.9
  ],
  "query_axis": [
    -1.2,
    -1.0,
    -0.8,
    -0.6,
    -0.4,
    -0.2,
    0.0,
    0.2,
    0.4,
    0.6,
    0.8,
    1.0,
    1.2
  ],
  "order": "y ascending, then x ascending",
  "synthetic_rule": "inside=1 if x*x+y*y < 0.36, otherwise outside=0",
  "conditions": {
    "kumo": [
      "original",
      "flip-00",
      "flip-01",
      "flip-02",
      "flip-03",
      "flip-04",
      "flip-05",
      "flip-06",
      "flip-07",
      "flip-08",
      "flip-09",
      "flip-10",
      "flip-11",
      "flip-12",
      "flip-13",
      "flip-14",
      "flip-15"
    ],
    "knn": [
      "original",
      "flip-00",
      "flip-01",
      "flip-02",
      "flip-03",
      "flip-04",
      "flip-05",
      "flip-06",
      "flip-07",
      "flip-08",
      "flip-09",
      "flip-10",
      "flip-11",
      "flip-12",
      "flip-13",
      "flip-14",
      "flip-15"
    ],
    "lr": [
      "original",
      "flip-00",
      "flip-01",
      "flip-02",
      "flip-03",
      "flip-04",
      "flip-05",
      "flip-06",
      "flip-07",
      "flip-08",
      "flip-09",
      "flip-10",
      "flip-11",
      "flip-12",
      "flip-13",
      "flip-14",
      "flip-15"
    ]
  },
  "settings": {
    "task": "classification",
    "size": "small",
    "device": "cpu",
    "dtype": "float32",
    "num_estimators": 1,
    "threads": 2,
    "recipe": "upstream default"
  },
  "report": [
    "all context/query inputs",
    "all 51 conditions including reused Kumo baseline",
    "native and named class probabilities",
    "per-condition timings and warnings/failures",
    "per-flip reach",
    "all-flip median/max summaries",
    "execution environment and resource receipt"
  ],
  "limits": {
    "inference_seconds": 180,
    "cpu_cores": 2,
    "memory_mib": 2048,
    "swap_mib": 0,
    "pids": 128,
    "network": "none during inference",
    "filesystem": "read-only source/dependencies/weights; one output directory and bounded tmpfs",
    "dollar_cost": 0,
    "downloads_max_mib": 1024,
    "abort": "timeout, memory limit, nonfinite output, wrong class order or device, missing or mismatched weight hash"
  },
  "claims": "One tiny synthetic two-feature table, one seed, one checkpoint, fixed model recipes. Preserve every condition and failure. No accuracy/calibration/general robustness claim; no scientific reproduction claimed.",
  "weight_sha256": "1ff91f484e19021aaf7d9eff6cc07e6d95a60b4aaad473e83e2ab2c3bfd2b617",
  "baseline": "Kumo original from results.json, retained byte-for-byte; no v1 inference rerun",
  "comparator_settings": {
    "scikit_learn": "1.9.1",
    "knn": "KNeighborsClassifier(n_neighbors=3), all other defaults",
    "lr": "LogisticRegression(), all defaults",
    "features": "identical float32 x/y arrays in original row order; no added scaling or feature engineering; upstream Kumo default recipe retained",
    "tie_rule": "argmax of [P(outside),P(inside)]; exact ties resolve to outside",
    "knn_ties": "default algorithm; equidistant neighbors can depend on fixed row order"
  },
  "reach_definition": {
    "changed_predictions": "argmax changes among all 169 fixed queries",
    "max_probability_change": "max absolute P(inside) shift among all 169 queries",
    "farthest_changed_distance": "Euclidean distance from flipped context row to farthest argmax-changed query, synthetic coordinate units; null when no classes change",
    "aggregate": "median and maximum across ALL 16 flips for each metric/model; no-change distance counts as 0 for aggregation"
  },
  "stop_test": {
    "failure": "STOP on any original abort condition or failed condition; retain partial output and missing conditions as not executed",
    "usefulness": "CONTINUE_PRIVATE_REVIEW only if Kumo vs either comparator differs by >=1 changed query in median/max or >=0.01 max P shift in median/max across all 16 flips; otherwise STOP",
    "interpretation": "private illustrative contrast only; not evidence of accuracy, calibration or superior robustness"
  },
  "comparator_source": "https://scikit-learn.org/stable/auto_examples/classification/plot_classifier_comparison.html"
}