{
 "manifest_version": 1,
 "role": "reference_line_tagger",
 "name": "reference-segmenter-geometry-tagger",
 "version": "sha256-fa3a2d07b46b",
 "task": "line_tagging",
 "runtime": {
  "engine": "onnxruntime",
  "requires": [
   "onnxruntime",
   "tokenizers",
   "numpy"
  ],
  "opset": 17,
  "executes_code_on_load": false
 },
 "files": [
  {
   "path": "gate.json",
   "bytes": 6963,
   "sha256": "3902d750ec2d8bb0dbd29e8079660f8411ff527b4a7d2d1e7bd7afb8a490acc6",
   "kind": "legacy_manifest"
  },
  {
   "path": "onnx/tagger.json",
   "bytes": 2290,
   "sha256": "3f3822418f7a6afd44bd86d280910deca288bfeaad06b3b7c21a65a8e9677167",
   "kind": "legacy_manifest"
  },
  {
   "path": "onnx/tagger.onnx",
   "bytes": 2410636,
   "sha256": "fa3a2d07b46bd9133fd7026ad09e502720a65413d5cdc1a736e9a9957526462b",
   "kind": "model"
  }
 ],
 "encoder": null,
 "labels": {
  "classes": [
   "continuation",
   "start"
  ]
 },
 "input": {
  "layout": "line_features",
  "tokenizer_file": null,
  "max_length": null,
  "features": [
   "caps_ratio",
   "contains_year",
   "dfont",
   "dx_prev",
   "dy_from_prev",
   "dy_gap",
   "dy_to_next",
   "early_year_paren",
   "font_size",
   "indent_delta_next",
   "indent_delta_prev",
   "is_first",
   "len_chars",
   "looks_author_year_start",
   "looks_doi_continuation",
   "looks_numbered_start",
   "next_len_chars",
   "next_looks_reference_header",
   "prev_ends_doi",
   "prev_ends_period",
   "prev_ends_year",
   "prev_len_chars",
   "prev_looks_reference_header",
   "same_page_next",
   "same_page_prev",
   "starts_author_date",
   "starts_numbered",
   "starts_year",
   "x0_rel_page"
  ],
  "window": null,
  "stride": null,
  "add_special_tokens": null,
  "truncation": null,
  "template": null,
  "normalization": null,
  "pair": null,
  "special_tokens": {},
  "model_inputs": [
   {
    "name": "features",
    "dtype": "float32",
    "shape": [
     "lines",
     29
    ]
   }
  ],
  "model_outputs": [
   {
    "name": "start_prob",
    "dtype": "float32",
    "shape": [
     "lines"
    ]
   },
   {
    "name": "logits",
    "dtype": "float32",
    "shape": [
     "lines",
     2
    ]
   }
  ]
 },
 "decode": {
  "method": "gated_line_starts",
  "crf": null,
  "calibration": {
   "file": "gate.json",
   "temperature": 1.5511548209195987
  },
  "when_nothing_found": "first_unit_forced",
  "thresholds": {
   "gate_confidence": 0.9666666666666667
  }
 },
 "quantization": {
  "weights": "fp32",
  "note": null
 },
 "provenance": {
  "ledger_key": "reference-segmenter-geometry-tagger",
  "run_record_sha256": "f0bcd8750673c7b64782b492e2bd8a26d10d4edac4d0312cecc9236ad370c03a",
  "platform_commit": "8c975b95c1a136957df82a5f72300ab11d1f29b1",
  "registry_build": "2026-11-10T21:38:59Z#d8d427de3ea7",
  "split_id": "corpus-split-v1@c6fc246571f4"
 },
 "licence": "mit",
 "compat": {
  "bibr_min_version": null,
  "export_schema": null
 },
 "self_test": {
  "tolerance": {
   "probability_abs": 0.01,
   "logit_abs": 0.05,
   "reason": "onnxruntime builds, CPUs and BLAS kernels give slightly different float results; the largest difference measured between onnxruntime 1.27.0 and 1.31.0 on the eight bundles normalised so far was 8e-05 (stored values are rounded to 5 places), so these limits are about two orders of magnitude looser than the drift seen, and far tighter than any change of a decision; tags and label lists are compared exactly"
  },
  "cases": [
   {
    "name": "six lines",
    "input": {
     "features": [
      [
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7
      ],
      [
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3
      ],
      [
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0
      ],
      [
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6
      ],
      [
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2
      ],
      [
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9,
       0.1,
       0.4,
       0.7,
       1.0,
       0.2,
       0.5,
       0.8,
       0.0,
       0.3,
       0.6,
       0.9
      ]
     ]
    },
    "expect": {
     "probs": [
      0.0044,
      0.10484,
      0.0029,
      0.06487,
      0.1367,
      0.00573
     ],
     "logits": [
      [
       4.14361,
       -4.26754
      ],
      [
       1.64345,
       -1.68308
      ]
     ]
    }
   }
  ]
 },
 "spec": {
  "gate_score": "expected_entry_precision",
  "layout": "lines first, no batch axis: one paper per call, features [lines, n] float32 in feature order"
 }
}