Publish case-context proxy rejection and source error audit

MANIFEST.json

@@ -1,11 +1,11 @@
 {
-  "schema": "aiflow-hf-research-snapshot-v27",
-  "generated_at": "2026-07-24T19:42:18.374253Z",
+  "schema": "aiflow-hf-research-snapshot-v28",
+  "generated_at": "2026-07-24T19:51:10.441196Z",
   "track": "R_noncommercial_plus_rejected_P_proxy",
   "product_validation": false,
   "public_release": true,
   "contains_raw_dataset": false,
-  "tests": "364 Python passed + 8 Android passed",
+  "tests": "367 Python passed + 8 Android passed",
   "retracted_paths": [
     "models/auxiliary/boundary_auxiliary_head.pt",
     "models/auxiliary/seed17/boundary_joint_delta.pt",
@@ -225,8 +225,8 @@
     },
     {
       "path": "README.md",
-      "bytes": 25440,
-      "sha256": "ed87f24c773f0839bcf7717e0f049ae4d8ea2a871f0a0541d64bab03babe967d"
+      "bytes": 25971,
+      "sha256": "e90002b2403aa23a4d8c525f7251bc098fcfff5520d73f5dd1b65f805b721afa"
     },
     {
       "path": "registry/math_ink_06_source_registry.json",
@@ -285,14 +285,19 @@
     },
     {
       "path": "reports/FEDERATION_AUDIT_20260724.md",
-      "bytes": 7085,
-      "sha256": "42a6f6897172c0e3962e0bf131d330c9b28b569e4d24e42db19c6075785f692b"
+      "bytes": 7634,
+      "sha256": "e7009ca51246d66f9ac823f9c988117897ce7e09137242902a8e124c7e2cbdbc"
     },
     {
       "path": "reports/federation_clean_seed17_error_audit.json",
       "bytes": 62353,
       "sha256": "6eb8552063155e0aa8885b575a7d8274c98ca77373ab6fd23694dc493782be20"
     },
+    {
+      "path": "reports/federation_clean_seed17_fullsource_error_audit.json",
+      "bytes": 63175,
+      "sha256": "e30732b319224c3d5c8b9a675e2ec135ddd6c2f8677555a62c1bf96fbcbafe85"
+    },
     {
       "path": "reports/federation_clean_seed17_fullsource_report.json",
       "bytes": 18173,
@@ -415,8 +420,8 @@
     },
     {
       "path": "scripts/audit_math_ink_06_federation_errors.py",
-      "bytes": 6752,
-      "sha256": "b03d6bb43a3d103ad19e205d2b29946a8dd93259097870607e4a6c7149459265"
+      "bytes": 10732,
+      "sha256": "bfe1e279a447d98cbc1c08fd42bf53ee06f7899b53710d7c6622fac58c64611a"
     },
     {
       "path": "scripts/audit_math_ink_06_formula_context_contract.py",

README.md

@@ -41,10 +41,13 @@ UJI Pen v1은 v2에 전부 포함된 mirror라 sampler에서 제거했다. HWRT
 
 이 결과는 출처 cap을 풀어도 UJI writer-generalization과 exact case family가 남는다는 근거다. Raster encoder·decoder만 여는 shadow run은 Pendigits를 개선했지만 HWRT holdout을 손상해 자동 guard에서 거부됐고 배포 모델에 포함하지 않는다. 3-seed distillation은 현재 보류 상태다.
 
+동일 writer의 고립 glyph를 baseline anchor로 쓰는 size-resolver proxy도 UJI에서 `51.09% → 45.63%`, HWRT에서 `53.57% → 39.29%`로 악화됐다. 이는 writer 묶음이 실제 수식 행 문맥이 아니라는 반증이다. 따라서 relative-size resolver는 현재 isolated symbol output에는 적용하지 않고, P Formula의 실제 행/Tray context가 확보된 경우에만 다시 검증한다.
+
 - [사람이 읽는 감사 보고서](reports/FEDERATION_AUDIT_20260724.md)
 - [기계 판독 federation audit](reports/federation_audit.json)
 - [교정 seed-17 full-source 결과](reports/federation_clean_seed17_fullsource_report.json)
 - [전체 오류/혼동 감사](reports/federation_clean_seed17_error_audit.json)
+- [full-source source별 오류·case proxy 감사](reports/federation_clean_seed17_fullsource_error_audit.json)
 - [출처 registry](registry/math_ink_06_source_registry.json)
 
 ## 핵심 입력 계약

reports/FEDERATION_AUDIT_20260724.md

@@ -98,4 +98,6 @@ UJI의 낮은 일반화가 2,000개 source cap 때문인지 확인하기 위해
 
 같은 checkpoint에서 raster encoder·causal virtual decoder만 2 epoch 학습한 shadow 실험도 수행했다. Pendigits raster가 validation 45.67%에서 최고 63.33%까지 오르지만 HWRT는 78.16%에서 71.65%로 떨어져 모든 interpolation이 source holdout guard에서 탈락했다. 이 변경은 checkpoint로 채택하지 않았고 3번째 epoch는 중단했다.
 
-다음 구현 우선순위는 (1) UJI의 case/O-0/cross family를 formula-relative size·neighbor context로 푸는 exact resolver, (2) source-aware virtual decoder 또는 더 다양한 paired raster↔stroke 데이터다. 이 두 작업 전에는 seed 31·47을 반복하지 않는다. 전체 Python 회귀는 364개를 통과했다.
+다음 구현 우선순위는 (1) UJI의 case/O-0/cross family를 formula-relative size·neighbor context로 푸는 exact resolver, (2) source-aware virtual decoder 또는 더 다양한 paired raster↔stroke 데이터다. 이 두 작업 전에는 seed 31·47을 반복하지 않는다. 전체 Python 회귀는 367개를 통과했다.
+
+상대크기 resolver를 현 UJI/HWRT test의 동일-writer isolated-glyph proxy에 적용한 결과는 UJI 51.09%→45.63%(-5.47%p), HWRT 53.57%→39.29%(-14.29%p)였다. 이 proxy는 실수식의 같은 행 baseline·상대크기가 아니라 writer의 서로 무관한 glyph를 anchor로 사용하므로 product evidence가 아니다. 다만 해로운 자동 뒤집힘을 실제로 확인했으므로 resolver는 현재 고립기호 0.6 반환에 연결하지 않고, 향후 P Formula row의 validated context가 있을 때만 activation한다.

reports/federation_clean_seed17_fullsource_error_audit.json

@@ -0,0 +1,2801 @@
+{
+  "samples": 7031,
+  "top1": 0.8412743806838989,
+  "top5": 0.9719812273979187,
+  "shape_family_top1": 0.8899160623550415,
+  "casefold_top1": 0.8805291056632996,
+  "case_only_error_rate": 0.24731182795698925,
+  "source_metrics": {
+    "hwrt": {
+      "samples": 552,
+      "top1": 0.8496376872062683,
+      "top5": 0.9909420013427734
+    },
+    "uci-pendigits": {
+      "samples": 3498,
+      "top1": 0.9268153309822083,
+      "top5": 0.9934248328208923
+    },
+    "uci-uji-pen-v2": {
+      "samples": 2981,
+      "top1": 0.7393491864204407,
+      "top5": 0.943307638168335
+    }
+  },
+  "writer_count": 49,
+  "writer_accuracy_p10": 0.581081081081081,
+  "writer_accuracy_minimum": 0.0,
+  "errors": 1116,
+  "same_family_error_rate": 0.3064516129032258,
+  "different_family_error_rate": 0.6935483870967742,
+  "top_confusions": [
+    {
+      "truth": "1",
+      "prediction": "2",
+      "count": 55
+    },
+    {
+      "truth": "7",
+      "prediction": "1",
+      "count": 41
+    },
+    {
+      "truth": "c",
+      "prediction": "C",
+      "count": 35
+    },
+    {
+      "truth": "s",
+      "prediction": "S",
+      "count": 30
+    },
+    {
+      "truth": "x",
+      "prediction": "X",
+      "count": 30
+    },
+    {
+      "truth": "o",
+      "prediction": "O",
+      "count": 17
+    },
+    {
+      "truth": "z",
+      "prediction": "Z",
+      "count": 17
+    },
+    {
+      "truth": "9",
+      "prediction": "3",
+      "count": 16
+    },
+    {
+      "truth": "5",
+      "prediction": "9",
+      "count": 15
+    },
+    {
+      "truth": "w",
+      "prediction": "W",
+      "count": 15
+    },
+    {
+      "truth": "P",
+      "prediction": "p",
+      "count": 14
+    },
+    {
+      "truth": "v",
+      "prediction": "V",
+      "count": 14
+    },
+    {
+      "truth": "W",
+      "prediction": "w",
+      "count": 14
+    },
+    {
+      "truth": "O",
+      "prediction": "0",
+      "count": 13
+    },
+    {
+      "truth": "2",
+      "prediction": "1",
+      "count": 12
+    },
+    {
+      "truth": "o",
+      "prediction": "0",
+      "count": 12
+    },
+    {
+      "truth": "3",
+      "prediction": "1",
+      "count": 11
+    },
+    {
+      "truth": "Z",
+      "prediction": "z",
+      "count": 11
+    },
+    {
+      "truth": "\"",
+      "prediction": "H",
+      "count": 11
+    },
+    {
+      "truth": "k",
+      "prediction": "K",
+      "count": 11
+    },
+    {
+      "truth": "3",
+      "prediction": "7",
+      "count": 10
+    },
+    {
+      "truth": "0",
+      "prediction": "O",
+      "count": 10
+    },
+    {
+      "truth": "5",
+      "prediction": "3",
+      "count": 9
+    },
+    {
+      "truth": "4",
+      "prediction": "5",
+      "count": 8
+    },
+    {
+      "truth": "8",
+      "prediction": "5",
+      "count": 8
+    },
+    {
+      "truth": ",",
+      "prediction": ")",
+      "count": 8
+    },
+    {
+      "truth": "\"",
+      "prediction": "\\parallel",
+      "count": 8
+    },
+    {
+      "truth": "p",
+      "prediction": "P",
+      "count": 8
+    },
+    {
+      "truth": "9",
+      "prediction": "q",
+      "count": 8
+    },
+    {
+      "truth": "7",
+      "prediction": "2",
+      "count": 7
+    }
+  ],
+  "highest_error_labels": [
+    {
+      "label": "1",
+      "test_support": 418,
+      "training_support": 908,
+      "top1": 0.8086124401913876,
+      "top5": 0.9904306220095693,
+      "errors": 80
+    },
+    {
+      "label": "7",
+      "test_support": 411,
+      "training_support": 904,
+      "top1": 0.8613138686131386,
+      "top5": 0.9708029197080292,
+      "errors": 57
+    },
+    {
+      "label": "9",
+      "test_support": 382,
+      "training_support": 861,
+      "top1": 0.8769633507853403,
+      "top5": 0.9895287958115183,
+      "errors": 47
+    },
+    {
+      "label": "x",
+      "test_support": 48,
+      "training_support": 120,
+      "top1": 0.10416666666666667,
+      "top5": 0.875,
+      "errors": 43
+    },
+    {
+      "label": "s",
+      "test_support": 47,
+      "training_support": 115,
+      "top1": 0.14893617021276595,
+      "top5": 0.9148936170212766,
+      "errors": 40
+    },
+    {
+      "label": "o",
+      "test_support": 47,
+      "training_support": 118,
+      "top1": 0.1702127659574468,
+      "top5": 0.9148936170212766,
+      "errors": 39
+    },
+    {
+      "label": "'",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.05,
+      "top5": 0.675,
+      "errors": 38
+    },
+    {
+      "label": "c",
+      "test_support": 45,
+      "training_support": 123,
+      "top1": 0.17777777777777778,
+      "top5": 1.0,
+      "errors": 37
+    },
+    {
+      "label": "5",
+      "test_support": 382,
+      "training_support": 852,
+      "top1": 0.9136125654450262,
+      "top5": 0.9947643979057592,
+      "errors": 33
+    },
+    {
+      "label": "\"",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.3,
+      "top5": 0.925,
+      "errors": 28
+    },
+    {
+      "label": "z",
+      "test_support": 47,
+      "training_support": 117,
+      "top1": 0.425531914893617,
+      "top5": 0.9148936170212766,
+      "errors": 27
+    },
+    {
+      "label": "8",
+      "test_support": 392,
+      "training_support": 852,
+      "top1": 0.9311224489795918,
+      "top5": 0.9974489795918368,
+      "errors": 27
+    },
+    {
+      "label": "v",
+      "test_support": 45,
+      "training_support": 121,
+      "top1": 0.4444444444444444,
+      "top5": 1.0,
+      "errors": 25
+    },
+    {
+      "label": "3",
+      "test_support": 381,
+      "training_support": 863,
+      "top1": 0.9396325459317585,
+      "top5": 0.9973753280839895,
+      "errors": 23
+    },
+    {
+      "label": "O",
+      "test_support": 45,
+      "training_support": 141,
+      "top1": 0.5111111111111111,
+      "top5": 0.9333333333333333,
+      "errors": 22
+    },
+    {
+      "label": ".",
+      "test_support": 22,
+      "training_support": 103,
+      "top1": 0.045454545454545456,
+      "top5": 0.3181818181818182,
+      "errors": 21
+    },
+    {
+      "label": "w",
+      "test_support": 45,
+      "training_support": 116,
+      "top1": 0.5333333333333333,
+      "top5": 0.9777777777777777,
+      "errors": 21
+    },
+    {
+      "label": "P",
+      "test_support": 47,
+      "training_support": 127,
+      "top1": 0.5531914893617021,
+      "top5": 0.9361702127659575,
+      "errors": 21
+    },
+    {
+      "label": "4",
+      "test_support": 410,
+      "training_support": 895,
+      "top1": 0.948780487804878,
+      "top5": 0.9951219512195122,
+      "errors": 21
+    },
+    {
+      "label": ",",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.5,
+      "top5": 0.65,
+      "errors": 20
+    },
+    {
+      "label": "Z",
+      "test_support": 44,
+      "training_support": 123,
+      "top1": 0.5909090909090909,
+      "top5": 0.9090909090909091,
+      "errors": 18
+    },
+    {
+      "label": "I",
+      "test_support": 49,
+      "training_support": 141,
+      "top1": 0.673469387755102,
+      "top5": 0.7959183673469388,
+      "errors": 16
+    },
+    {
+      "label": "2",
+      "test_support": 414,
+      "training_support": 915,
+      "top1": 0.961352657004831,
+      "top5": 0.9927536231884058,
+      "errors": 16
+    },
+    {
+      "label": "k",
+      "test_support": 43,
+      "training_support": 119,
+      "top1": 0.6511627906976745,
+      "top5": 0.9767441860465116,
+      "errors": 15
+    },
+    {
+      "label": "q",
+      "test_support": 45,
+      "training_support": 128,
+      "top1": 0.6666666666666666,
+      "top5": 0.9333333333333333,
+      "errors": 15
+    },
+    {
+      "label": "y",
+      "test_support": 45,
+      "training_support": 118,
+      "top1": 0.6666666666666666,
+      "top5": 0.9333333333333333,
+      "errors": 15
+    },
+    {
+      "label": "p",
+      "test_support": 47,
+      "training_support": 113,
+      "top1": 0.6808510638297872,
+      "top5": 0.9574468085106383,
+      "errors": 15
+    },
+    {
+      "label": "V",
+      "test_support": 43,
+      "training_support": 119,
+      "top1": 0.6744186046511628,
+      "top5": 0.9534883720930233,
+      "errors": 14
+    },
+    {
+      "label": "W",
+      "test_support": 44,
+      "training_support": 122,
+      "top1": 0.6818181818181818,
+      "top5": 1.0,
+      "errors": 14
+    },
+    {
+      "label": "S",
+      "test_support": 45,
+      "training_support": 122,
+      "top1": 0.6888888888888889,
+      "top5": 0.9333333333333333,
+      "errors": 14
+    },
+    {
+      "label": "0",
+      "test_support": 406,
+      "training_support": 927,
+      "top1": 0.9655172413793104,
+      "top5": 1.0,
+      "errors": 14
+    },
+    {
+      "label": "t",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.675,
+      "top5": 0.9,
+      "errors": 13
+    },
+    {
+      "label": "U",
+      "test_support": 45,
+      "training_support": 118,
+      "top1": 0.7111111111111111,
+      "top5": 1.0,
+      "errors": 13
+    },
+    {
+      "label": "K",
+      "test_support": 50,
+      "training_support": 141,
+      "top1": 0.76,
+      "top5": 0.92,
+      "errors": 12
+    },
+    {
+      "label": "Y",
+      "test_support": 43,
+      "training_support": 117,
+      "top1": 0.7441860465116279,
+      "top5": 0.8837209302325582,
+      "errors": 11
+    },
+    {
+      "label": "M",
+      "test_support": 44,
+      "training_support": 135,
+      "top1": 0.75,
+      "top5": 0.9772727272727273,
+      "errors": 11
+    },
+    {
+      "label": "X",
+      "test_support": 45,
+      "training_support": 113,
+      "top1": 0.7555555555555555,
+      "top5": 0.9333333333333333,
+      "errors": 11
+    },
+    {
+      "label": "u",
+      "test_support": 46,
+      "training_support": 115,
+      "top1": 0.7608695652173914,
+      "top5": 0.9347826086956522,
+      "errors": 11
+    },
+    {
+      "label": "T",
+      "test_support": 45,
+      "training_support": 114,
+      "top1": 0.7777777777777778,
+      "top5": 0.9111111111111111,
+      "errors": 10
+    },
+    {
+      "label": "J",
+      "test_support": 44,
+      "training_support": 145,
+      "top1": 0.7954545454545454,
+      "top5": 0.9772727272727273,
+      "errors": 9
+    }
+  ],
+  "lowest_supported_labels": [
+    {
+      "label": "\\mathbb{N}",
+      "test_support": 2,
+      "training_support": 41,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\mathbb{R}",
+      "test_support": 4,
+      "training_support": 44,
+      "top1": 0.75,
+      "top5": 1.0,
+      "errors": 1
+    },
+    {
+      "label": "\\mathbb{Q}",
+      "test_support": 1,
+      "training_support": 44,
+      "top1": 0.0,
+      "top5": 1.0,
+      "errors": 1
+    },
+    {
+      "label": "\\blacktriangleright",
+      "test_support": 1,
+      "training_support": 52,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\mathbb{Z}",
+      "test_support": 2,
+      "training_support": 56,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "+",
+      "test_support": 16,
+      "training_support": 58,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\mathbb{H}",
+      "test_support": 2,
+      "training_support": 61,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "!",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.925,
+      "top5": 1.0,
+      "errors": 3
+    },
+    {
+      "label": "\"",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.3,
+      "top5": 0.925,
+      "errors": 28
+    },
+    {
+      "label": "'",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.05,
+      "top5": 0.675,
+      "errors": 38
+    },
+    {
+      "label": "(",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.825,
+      "top5": 1.0,
+      "errors": 7
+    },
+    {
+      "label": ")",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.925,
+      "top5": 1.0,
+      "errors": 3
+    },
+    {
+      "label": ",",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.5,
+      "top5": 0.65,
+      "errors": 20
+    },
+    {
+      "label": ":",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.825,
+      "top5": 1.0,
+      "errors": 7
+    },
+    {
+      "label": ";",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.825,
+      "top5": 0.975,
+      "errors": 7
+    },
+    {
+      "label": "t",
+      "test_support": 40,
+      "training_support": 72,
+      "top1": 0.675,
+      "top5": 0.9,
+      "errors": 13
+    },
+    {
+      "label": "\\subset",
+      "test_support": 4,
+      "training_support": 75,
+      "top1": 0.25,
+      "top5": 1.0,
+      "errors": 3
+    },
+    {
+      "label": "\\barwedge",
+      "test_support": 3,
+      "training_support": 75,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\sqcap",
+      "test_support": 4,
+      "training_support": 76,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\phi",
+      "test_support": 3,
+      "training_support": 76,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\Gamma",
+      "test_support": 1,
+      "training_support": 76,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\amalg",
+      "test_support": 3,
+      "training_support": 77,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\kappa",
+      "test_support": 3,
+      "training_support": 77,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\mathcal{G}",
+      "test_support": 3,
+      "training_support": 77,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\mathfrak{A}",
+      "test_support": 3,
+      "training_support": 77,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\mathfrak{X}",
+      "test_support": 3,
+      "training_support": 77,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\rfloor",
+      "test_support": 3,
+      "training_support": 77,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\varkappa",
+      "test_support": 3,
+      "training_support": 77,
+      "top1": 0.0,
+      "top5": 0.6666666666666666,
+      "errors": 3
+    },
+    {
+      "label": "\\iota",
+      "test_support": 2,
+      "training_support": 77,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\parr",
+      "test_support": 2,
+      "training_support": 77,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\varrho",
+      "test_support": 2,
+      "training_support": 77,
+      "top1": 0.5,
+      "top5": 1.0,
+      "errors": 1
+    },
+    {
+      "label": "\\Theta",
+      "test_support": 2,
+      "training_support": 78,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\Xi",
+      "test_support": 2,
+      "training_support": 78,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\circledcirc",
+      "test_support": 2,
+      "training_support": 78,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\dotsc",
+      "test_support": 2,
+      "training_support": 78,
+      "top1": 0.5,
+      "top5": 1.0,
+      "errors": 1
+    },
+    {
+      "label": "\\epsilon",
+      "test_support": 2,
+      "training_support": 78,
+      "top1": 0.0,
+      "top5": 1.0,
+      "errors": 2
+    },
+    {
+      "label": "\\oiint",
+      "test_support": 2,
+      "training_support": 78,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\pitchfork",
+      "test_support": 2,
+      "training_support": 78,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\zeta",
+      "test_support": 2,
+      "training_support": 78,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    },
+    {
+      "label": "\\|",
+      "test_support": 2,
+      "training_support": 78,
+      "top1": 1.0,
+      "top5": 1.0,
+      "errors": 0
+    }
+  ],
+  "source_bottlenecks": {
+    "hwrt": {
+      "samples": 552,
+      "top1": 0.8496376872062683,
+      "top5": 0.9909420013427734,
+      "shape_family_top1": 0.9221014380455017,
+      "same_family_error_rate": 0.4819277108433735,
+      "writer_accuracy_p10": 0.7105263157894737,
+      "writer_accuracy_minimum": 0.0,
+      "top_confusions": [
+        {
+          "truth": "1",
+          "prediction": "I",
+          "count": 6
+        },
+        {
+          "truth": "o",
+          "prediction": "0",
+          "count": 4
+        },
+        {
+          "truth": "O",
+          "prediction": "0",
+          "count": 4
+        },
+        {
+          "truth": "v",
+          "prediction": "V",
+          "count": 4
+        },
+        {
+          "truth": "s",
+          "prediction": "S",
+          "count": 4
+        },
+        {
+          "truth": "z",
+          "prediction": "Z",
+          "count": 4
+        },
+        {
+          "truth": "x",
+          "prediction": "X",
+          "count": 3
+        },
+        {
+          "truth": "\\subset",
+          "prediction": "c",
+          "count": 3
+        },
+        {
+          "truth": "9",
+          "prediction": "g",
+          "count": 3
+        },
+        {
+          "truth": "X",
+          "prediction": "x",
+          "count": 2
+        },
+        {
+          "truth": "\\varkappa",
+          "prediction": "\\kappa",
+          "count": 2
+        },
+        {
+          "truth": "w",
+          "prediction": "W",
+          "count": 2
+        },
+        {
+          "truth": "c",
+          "prediction": "C",
+          "count": 1
+        },
+        {
+          "truth": "0",
+          "prediction": "O",
+          "count": 1
+        },
+        {
+          "truth": "O",
+          "prediction": "\\fullmoon",
+          "count": 1
+        },
+        {
+          "truth": "\\dotsc",
+          "prediction": "-",
+          "count": 1
+        },
+        {
+          "truth": "x",
+          "prediction": "\\ae",
+          "count": 1
+        },
+        {
+          "truth": "l",
+          "prediction": "I",
+          "count": 1
+        },
+        {
+          "truth": "l",
+          "prediction": "'",
+          "count": 1
+        },
+        {
+          "truth": "\\mathbb{R}",
+          "prediction": "\\mathds{R}",
+          "count": 1
+        },
+        {
+          "truth": "\\circ",
+          "prediction": "0",
+          "count": 1
+        },
+        {
+          "truth": "I",
+          "prediction": "|",
+          "count": 1
+        },
+        {
+          "truth": "I",
+          "prediction": "\\mid",
+          "count": 1
+        },
+        {
+          "truth": "\\&",
+          "prediction": "\\with",
+          "count": 1
+        },
+        {
+          "truth": "\\varrho",
+          "prediction": "\\rho",
+          "count": 1
+        },
+        {
+          "truth": "U",
+          "prediction": "\\iota",
+          "count": 1
+        },
+        {
+          "truth": "o",
+          "prediction": "\\fullmoon",
+          "count": 1
+        },
+        {
+          "truth": "\\dots",
+          "prediction": "-",
+          "count": 1
+        },
+        {
+          "truth": "9",
+          "prediction": "0",
+          "count": 1
+        },
+        {
+          "truth": "x",
+          "prediction": "\\chi",
+          "count": 1
+        }
+      ],
+      "highest_error_labels": [
+        {
+          "label": "o",
+          "test_support": 7,
+          "training_support": 46,
+          "top1": 0.14285714285714285,
+          "top5": 0.8571428571428571,
+          "errors": 6
+        },
+        {
+          "label": "x",
+          "test_support": 8,
+          "training_support": 48,
+          "top1": 0.25,
+          "top5": 0.875,
+          "errors": 6
+        },
+        {
+          "label": "1",
+          "test_support": 14,
+          "training_support": 57,
+          "top1": 0.5714285714285714,
+          "top5": 1.0,
+          "errors": 6
+        },
+        {
+          "label": "O",
+          "test_support": 5,
+          "training_support": 69,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 5
+        },
+        {
+          "label": "z",
+          "test_support": 7,
+          "training_support": 45,
+          "top1": 0.2857142857142857,
+          "top5": 1.0,
+          "errors": 5
+        },
+        {
+          "label": "l",
+          "test_support": 5,
+          "training_support": 43,
+          "top1": 0.2,
+          "top5": 0.8,
+          "errors": 4
+        },
+        {
+          "label": "v",
+          "test_support": 5,
+          "training_support": 49,
+          "top1": 0.2,
+          "top5": 1.0,
+          "errors": 4
+        },
+        {
+          "label": "9",
+          "test_support": 6,
+          "training_support": 70,
+          "top1": 0.3333333333333333,
+          "top5": 1.0,
+          "errors": 4
+        },
+        {
+          "label": "s",
+          "test_support": 7,
+          "training_support": 43,
+          "top1": 0.42857142857142855,
+          "top5": 1.0,
+          "errors": 4
+        },
+        {
+          "label": "\\varkappa",
+          "test_support": 3,
+          "training_support": 77,
+          "top1": 0.0,
+          "top5": 0.6666666666666666,
+          "errors": 3
+        },
+        {
+          "label": "\\subset",
+          "test_support": 4,
+          "training_support": 75,
+          "top1": 0.25,
+          "top5": 1.0,
+          "errors": 3
+        },
+        {
+          "label": "X",
+          "test_support": 5,
+          "training_support": 41,
+          "top1": 0.4,
+          "top5": 1.0,
+          "errors": 3
+        },
+        {
+          "label": "\\epsilon",
+          "test_support": 2,
+          "training_support": 78,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 2
+        },
+        {
+          "label": "w",
+          "test_support": 5,
+          "training_support": 44,
+          "top1": 0.6,
+          "top5": 1.0,
+          "errors": 2
+        },
+        {
+          "label": "I",
+          "test_support": 9,
+          "training_support": 69,
+          "top1": 0.7777777777777778,
+          "top5": 1.0,
+          "errors": 2
+        },
+        {
+          "label": ".",
+          "test_support": 1,
+          "training_support": 59,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\&",
+          "test_support": 1,
+          "training_support": 79,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\circ",
+          "test_support": 1,
+          "training_support": 79,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\dots",
+          "test_support": 1,
+          "training_support": 78,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\ell",
+          "test_support": 1,
+          "training_support": 79,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\fint",
+          "test_support": 1,
+          "training_support": 79,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\mathbb{Q}",
+          "test_support": 1,
+          "training_support": 44,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\mid",
+          "test_support": 1,
+          "training_support": 79,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\sqrt{}",
+          "test_support": 1,
+          "training_support": 79,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\vee",
+          "test_support": 1,
+          "training_support": 79,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "|",
+          "test_support": 1,
+          "training_support": 79,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\dotsc",
+          "test_support": 2,
+          "training_support": 78,
+          "top1": 0.5,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\varrho",
+          "test_support": 2,
+          "training_support": 77,
+          "top1": 0.5,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "0",
+          "test_support": 3,
+          "training_support": 75,
+          "top1": 0.6666666666666666,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "\\mathbb{R}",
+          "test_support": 4,
+          "training_support": 44,
+          "top1": 0.75,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "S",
+          "test_support": 5,
+          "training_support": 50,
+          "top1": 0.8,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "U",
+          "test_support": 5,
+          "training_support": 46,
+          "top1": 0.8,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "c",
+          "test_support": 5,
+          "training_support": 51,
+          "top1": 0.8,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "i",
+          "test_support": 6,
+          "training_support": 44,
+          "top1": 0.8333333333333334,
+          "top5": 0.8333333333333334,
+          "errors": 1
+        },
+        {
+          "label": "u",
+          "test_support": 6,
+          "training_support": 43,
+          "top1": 0.8333333333333334,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "5",
+          "test_support": 7,
+          "training_support": 60,
+          "top1": 0.8571428571428571,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "P",
+          "test_support": 7,
+          "training_support": 55,
+          "top1": 0.8571428571428571,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "p",
+          "test_support": 7,
+          "training_support": 41,
+          "top1": 0.8571428571428571,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "a",
+          "test_support": 9,
+          "training_support": 67,
+          "top1": 0.8888888888888888,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "%",
+          "test_support": 1,
+          "training_support": 79,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        }
+      ],
+      "lowest_supported_labels": [
+        {
+          "label": "j",
+          "test_support": 7,
+          "training_support": 37,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "B",
+          "test_support": 3,
+          "training_support": 37,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "D",
+          "test_support": 9,
+          "training_support": 39,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "b",
+          "test_support": 8,
+          "training_support": 41,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "g",
+          "test_support": 7,
+          "training_support": 41,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "p",
+          "test_support": 7,
+          "training_support": 41,
+          "top1": 0.8571428571428571,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "X",
+          "test_support": 5,
+          "training_support": 41,
+          "top1": 0.4,
+          "top5": 1.0,
+          "errors": 3
+        },
+        {
+          "label": "\\mathbb{N}",
+          "test_support": 2,
+          "training_support": 41,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "A",
+          "test_support": 10,
+          "training_support": 42,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "m",
+          "test_support": 7,
+          "training_support": 42,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "T",
+          "test_support": 5,
+          "training_support": 42,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "s",
+          "test_support": 7,
+          "training_support": 43,
+          "top1": 0.42857142857142855,
+          "top5": 1.0,
+          "errors": 4
+        },
+        {
+          "label": "4",
+          "test_support": 6,
+          "training_support": 43,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "u",
+          "test_support": 6,
+          "training_support": 43,
+          "top1": 0.8333333333333334,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "l",
+          "test_support": 5,
+          "training_support": 43,
+          "top1": 0.2,
+          "top5": 0.8,
+          "errors": 4
+        },
+        {
+          "label": "e",
+          "test_support": 10,
+          "training_support": 44,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "n",
+          "test_support": 7,
+          "training_support": 44,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "i",
+          "test_support": 6,
+          "training_support": 44,
+          "top1": 0.8333333333333334,
+          "top5": 0.8333333333333334,
+          "errors": 1
+        },
+        {
+          "label": "F",
+          "test_support": 5,
+          "training_support": 44,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "w",
+          "test_support": 5,
+          "training_support": 44,
+          "top1": 0.6,
+          "top5": 1.0,
+          "errors": 2
+        },
+        {
+          "label": "\\mathbb{R}",
+          "test_support": 4,
+          "training_support": 44,
+          "top1": 0.75,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "h",
+          "test_support": 4,
+          "training_support": 44,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "E",
+          "test_support": 3,
+          "training_support": 44,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "\\mathbb{Q}",
+          "test_support": 1,
+          "training_support": 44,
+          "top1": 0.0,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "z",
+          "test_support": 7,
+          "training_support": 45,
+          "top1": 0.2857142857142857,
+          "top5": 1.0,
+          "errors": 5
+        },
+        {
+          "label": "d",
+          "test_support": 5,
+          "training_support": 45,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "Y",
+          "test_support": 3,
+          "training_support": 45,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "o",
+          "test_support": 7,
+          "training_support": 46,
+          "top1": 0.14285714285714285,
+          "top5": 0.8571428571428571,
+          "errors": 6
+        },
+        {
+          "label": "U",
+          "test_support": 5,
+          "training_support": 46,
+          "top1": 0.8,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "r",
+          "test_support": 5,
+          "training_support": 46,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "y",
+          "test_support": 5,
+          "training_support": 46,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "V",
+          "test_support": 3,
+          "training_support": 47,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "k",
+          "test_support": 3,
+          "training_support": 47,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "f",
+          "test_support": 8,
+          "training_support": 48,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "x",
+          "test_support": 8,
+          "training_support": 48,
+          "top1": 0.25,
+          "top5": 0.875,
+          "errors": 6
+        },
+        {
+          "label": "v",
+          "test_support": 5,
+          "training_support": 49,
+          "top1": 0.2,
+          "top5": 1.0,
+          "errors": 4
+        },
+        {
+          "label": "S",
+          "test_support": 5,
+          "training_support": 50,
+          "top1": 0.8,
+          "top5": 1.0,
+          "errors": 1
+        },
+        {
+          "label": "W",
+          "test_support": 4,
+          "training_support": 50,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "H",
+          "test_support": 3,
+          "training_support": 50,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "c",
+          "test_support": 5,
+          "training_support": 51,
+          "top1": 0.8,
+          "top5": 1.0,
+          "errors": 1
+        }
+      ]
+    },
+    "uci-pendigits": {
+      "samples": 3498,
+      "top1": 0.9268153309822083,
+      "top5": 0.9934248328208923,
+      "shape_family_top1": 0.9279588460922241,
+      "same_family_error_rate": 0.015625,
+      "writer_accuracy_p10": 0.9268153230417381,
+      "writer_accuracy_minimum": 0.9268153230417381,
+      "top_confusions": [
+        {
+          "truth": "1",
+          "prediction": "2",
+          "count": 55
+        },
+        {
+          "truth": "7",
+          "prediction": "1",
+          "count": 40
+        },
+        {
+          "truth": "9",
+          "prediction": "3",
+          "count": 16
+        },
+        {
+          "truth": "5",
+          "prediction": "9",
+          "count": 15
+        },
+        {
+          "truth": "2",
+          "prediction": "1",
+          "count": 12
+        },
+        {
+          "truth": "3",
+          "prediction": "1",
+          "count": 11
+        },
+        {
+          "truth": "3",
+          "prediction": "7",
+          "count": 10
+        },
+        {
+          "truth": "5",
+          "prediction": "3",
+          "count": 9
+        },
+        {
+          "truth": "4",
+          "prediction": "5",
+          "count": 8
+        },
+        {
+          "truth": "8",
+          "prediction": "5",
+          "count": 8
+        },
+        {
+          "truth": "8",
+          "prediction": "9",
+          "count": 7
+        },
+        {
+          "truth": "9",
+          "prediction": "\\partial",
+          "count": 7
+        },
+        {
+          "truth": "7",
+          "prediction": "2",
+          "count": 6
+        },
+        {
+          "truth": "6",
+          "prediction": "1",
+          "count": 4
+        },
+        {
+          "truth": "1",
+          "prediction": "L",
+          "count": 4
+        },
+        {
+          "truth": "8",
+          "prediction": "4",
+          "count": 3
+        },
+        {
+          "truth": "4",
+          "prediction": "6",
+          "count": 3
+        },
+        {
+          "truth": "8",
+          "prediction": "\\triangleright",
+          "count": 2
+        },
+        {
+          "truth": "5",
+          "prediction": "e",
+          "count": 2
+        },
+        {
+          "truth": "8",
+          "prediction": "0",
+          "count": 2
+        },
+        {
+          "truth": "7",
+          "prediction": "\\rho",
+          "count": 2
+        },
+        {
+          "truth": "9",
+          "prediction": "8",
+          "count": 2
+        },
+        {
+          "truth": "7",
+          "prediction": "8",
+          "count": 2
+        },
+        {
+          "truth": "9",
+          "prediction": "1",
+          "count": 2
+        },
+        {
+          "truth": "5",
+          "prediction": "8",
+          "count": 2
+        },
+        {
+          "truth": "0",
+          "prediction": "\\circlearrowright",
+          "count": 2
+        },
+        {
+          "truth": "3",
+          "prediction": "5",
+          "count": 1
+        },
+        {
+          "truth": "8",
+          "prediction": "6",
+          "count": 1
+        },
+        {
+          "truth": "1",
+          "prediction": "r",
+          "count": 1
+        },
+        {
+          "truth": "2",
+          "prediction": "6",
+          "count": 1
+        }
+      ],
+      "highest_error_labels": [
+        {
+          "label": "1",
+          "test_support": 364,
+          "training_support": 779,
+          "top1": 0.8214285714285714,
+          "top5": 0.989010989010989,
+          "errors": 65
+        },
+        {
+          "label": "7",
+          "test_support": 364,
+          "training_support": 778,
+          "top1": 0.8598901098901099,
+          "top5": 0.9752747252747253,
+          "errors": 51
+        },
+        {
+          "label": "9",
+          "test_support": 336,
+          "training_support": 719,
+          "top1": 0.9107142857142857,
+          "top5": 0.9910714285714286,
+          "errors": 30
+        },
+        {
+          "label": "5",
+          "test_support": 335,
+          "training_support": 720,
+          "top1": 0.9134328358208955,
+          "top5": 0.9970149253731343,
+          "errors": 29
+        },
+        {
+          "label": "8",
+          "test_support": 336,
+          "training_support": 719,
+          "top1": 0.9255952380952381,
+          "top5": 0.9970238095238095,
+          "errors": 25
+        },
+        {
+          "label": "3",
+          "test_support": 336,
+          "training_support": 719,
+          "top1": 0.9315476190476191,
+          "top5": 0.9970238095238095,
+          "errors": 23
+        },
+        {
+          "label": "2",
+          "test_support": 364,
+          "training_support": 780,
+          "top1": 0.9615384615384616,
+          "top5": 0.9917582417582418,
+          "errors": 14
+        },
+        {
+          "label": "4",
+          "test_support": 364,
+          "training_support": 780,
+          "top1": 0.967032967032967,
+          "top5": 0.9972527472527473,
+          "errors": 12
+        },
+        {
+          "label": "6",
+          "test_support": 336,
+          "training_support": 720,
+          "top1": 0.9851190476190477,
+          "top5": 1.0,
+          "errors": 5
+        },
+        {
+          "label": "0",
+          "test_support": 363,
+          "training_support": 780,
+          "top1": 0.9944903581267218,
+          "top5": 1.0,
+          "errors": 2
+        }
+      ],
+      "lowest_supported_labels": [
+        {
+          "label": "3",
+          "test_support": 336,
+          "training_support": 719,
+          "top1": 0.9315476190476191,
+          "top5": 0.9970238095238095,
+          "errors": 23
+        },
+        {
+          "label": "8",
+          "test_support": 336,
+          "training_support": 719,
+          "top1": 0.9255952380952381,
+          "top5": 0.9970238095238095,
+          "errors": 25
+        },
+        {
+          "label": "9",
+          "test_support": 336,
+          "training_support": 719,
+          "top1": 0.9107142857142857,
+          "top5": 0.9910714285714286,
+          "errors": 30
+        },
+        {
+          "label": "6",
+          "test_support": 336,
+          "training_support": 720,
+          "top1": 0.9851190476190477,
+          "top5": 1.0,
+          "errors": 5
+        },
+        {
+          "label": "5",
+          "test_support": 335,
+          "training_support": 720,
+          "top1": 0.9134328358208955,
+          "top5": 0.9970149253731343,
+          "errors": 29
+        },
+        {
+          "label": "7",
+          "test_support": 364,
+          "training_support": 778,
+          "top1": 0.8598901098901099,
+          "top5": 0.9752747252747253,
+          "errors": 51
+        },
+        {
+          "label": "1",
+          "test_support": 364,
+          "training_support": 779,
+          "top1": 0.8214285714285714,
+          "top5": 0.989010989010989,
+          "errors": 65
+        },
+        {
+          "label": "2",
+          "test_support": 364,
+          "training_support": 780,
+          "top1": 0.9615384615384616,
+          "top5": 0.9917582417582418,
+          "errors": 14
+        },
+        {
+          "label": "4",
+          "test_support": 364,
+          "training_support": 780,
+          "top1": 0.967032967032967,
+          "top5": 0.9972527472527473,
+          "errors": 12
+        },
+        {
+          "label": "0",
+          "test_support": 363,
+          "training_support": 780,
+          "top1": 0.9944903581267218,
+          "top5": 1.0,
+          "errors": 2
+        }
+      ]
+    },
+    "uci-uji-pen-v2": {
+      "samples": 2981,
+      "top1": 0.7393491864204407,
+      "top5": 0.943307638168335,
+      "shape_family_top1": 0.83931565284729,
+      "same_family_error_rate": 0.3835263835263835,
+      "writer_accuracy_p10": 0.5666666666666667,
+      "writer_accuracy_minimum": 0.5266666666666666,
+      "top_confusions": [
+        {
+          "truth": "c",
+          "prediction": "C",
+          "count": 34
+        },
+        {
+          "truth": "x",
+          "prediction": "X",
+          "count": 27
+        },
+        {
+          "truth": "s",
+          "prediction": "S",
+          "count": 26
+        },
+        {
+          "truth": "o",
+          "prediction": "O",
+          "count": 16
+        },
+        {
+          "truth": "P",
+          "prediction": "p",
+          "count": 14
+        },
+        {
+          "truth": "W",
+          "prediction": "w",
+          "count": 14
+        },
+        {
+          "truth": "w",
+          "prediction": "W",
+          "count": 13
+        },
+        {
+          "truth": "z",
+          "prediction": "Z",
+          "count": 13
+        },
+        {
+          "truth": "Z",
+          "prediction": "z",
+          "count": 11
+        },
+        {
+          "truth": "\"",
+          "prediction": "H",
+          "count": 11
+        },
+        {
+          "truth": "k",
+          "prediction": "K",
+          "count": 11
+        },
+        {
+          "truth": "v",
+          "prediction": "V",
+          "count": 10
+        },
+        {
+          "truth": "O",
+          "prediction": "0",
+          "count": 9
+        },
+        {
+          "truth": "0",
+          "prediction": "O",
+          "count": 9
+        },
+        {
+          "truth": ",",
+          "prediction": ")",
+          "count": 8
+        },
+        {
+          "truth": "\"",
+          "prediction": "\\parallel",
+          "count": 8
+        },
+        {
+          "truth": "o",
+          "prediction": "0",
+          "count": 8
+        },
+        {
+          "truth": "9",
+          "prediction": "q",
+          "count": 8
+        },
+        {
+          "truth": "(",
+          "prediction": "C",
+          "count": 7
+        },
+        {
+          "truth": "'",
+          "prediction": "I",
+          "count": 7
+        },
+        {
+          "truth": "p",
+          "prediction": "P",
+          "count": 7
+        },
+        {
+          "truth": "M",
+          "prediction": "m",
+          "count": 6
+        },
+        {
+          "truth": "Y",
+          "prediction": "y",
+          "count": 6
+        },
+        {
+          "truth": "1",
+          "prediction": "l",
+          "count": 6
+        },
+        {
+          "truth": "'",
+          "prediction": ",",
+          "count": 6
+        },
+        {
+          "truth": "K",
+          "prediction": "k",
+          "count": 6
+        },
+        {
+          "truth": "y",
+          "prediction": "Y",
+          "count": 6
+        },
+        {
+          "truth": "q",
+          "prediction": "9",
+          "count": 5
+        },
+        {
+          "truth": "w",
+          "prediction": "\\omega",
+          "count": 5
+        },
+        {
+          "truth": "L",
+          "prediction": "\\mathcal{L}",
+          "count": 5
+        }
+      ],
+      "highest_error_labels": [
+        {
+          "label": "'",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.05,
+          "top5": 0.675,
+          "errors": 38
+        },
+        {
+          "label": "x",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.075,
+          "top5": 0.875,
+          "errors": 37
+        },
+        {
+          "label": "c",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.1,
+          "top5": 1.0,
+          "errors": 36
+        },
+        {
+          "label": "s",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.1,
+          "top5": 0.9,
+          "errors": 36
+        },
+        {
+          "label": "o",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.175,
+          "top5": 0.925,
+          "errors": 33
+        },
+        {
+          "label": "\"",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.3,
+          "top5": 0.925,
+          "errors": 28
+        },
+        {
+          "label": "z",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.45,
+          "top5": 0.9,
+          "errors": 22
+        },
+        {
+          "label": "v",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.475,
+          "top5": 1.0,
+          "errors": 21
+        },
+        {
+          "label": ".",
+          "test_support": 21,
+          "training_support": 44,
+          "top1": 0.047619047619047616,
+          "top5": 0.2857142857142857,
+          "errors": 20
+        },
+        {
+          "label": ",",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.5,
+          "top5": 0.65,
+          "errors": 20
+        },
+        {
+          "label": "P",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.5,
+          "top5": 0.925,
+          "errors": 20
+        },
+        {
+          "label": "w",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.525,
+          "top5": 0.975,
+          "errors": 19
+        },
+        {
+          "label": "Z",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.55,
+          "top5": 0.9,
+          "errors": 18
+        },
+        {
+          "label": "O",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.575,
+          "top5": 0.925,
+          "errors": 17
+        },
+        {
+          "label": "k",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.625,
+          "top5": 0.975,
+          "errors": 15
+        },
+        {
+          "label": "q",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.625,
+          "top5": 0.925,
+          "errors": 15
+        },
+        {
+          "label": "y",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.625,
+          "top5": 0.925,
+          "errors": 15
+        },
+        {
+          "label": "I",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.65,
+          "top5": 0.75,
+          "errors": 14
+        },
+        {
+          "label": "V",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.65,
+          "top5": 0.95,
+          "errors": 14
+        },
+        {
+          "label": "W",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.65,
+          "top5": 1.0,
+          "errors": 14
+        },
+        {
+          "label": "p",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.65,
+          "top5": 0.95,
+          "errors": 14
+        },
+        {
+          "label": "9",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.675,
+          "top5": 0.975,
+          "errors": 13
+        },
+        {
+          "label": "S",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.675,
+          "top5": 0.925,
+          "errors": 13
+        },
+        {
+          "label": "t",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.675,
+          "top5": 0.9,
+          "errors": 13
+        },
+        {
+          "label": "K",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.7,
+          "top5": 0.9,
+          "errors": 12
+        },
+        {
+          "label": "U",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.7,
+          "top5": 1.0,
+          "errors": 12
+        },
+        {
+          "label": "0",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.725,
+          "top5": 1.0,
+          "errors": 11
+        },
+        {
+          "label": "M",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.725,
+          "top5": 0.975,
+          "errors": 11
+        },
+        {
+          "label": "Y",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.725,
+          "top5": 0.875,
+          "errors": 11
+        },
+        {
+          "label": "T",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.75,
+          "top5": 0.9,
+          "errors": 10
+        },
+        {
+          "label": "u",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.75,
+          "top5": 0.925,
+          "errors": 10
+        },
+        {
+          "label": "1",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.775,
+          "top5": 1.0,
+          "errors": 9
+        },
+        {
+          "label": "4",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.775,
+          "top5": 0.975,
+          "errors": 9
+        },
+        {
+          "label": "J",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.775,
+          "top5": 0.975,
+          "errors": 9
+        },
+        {
+          "label": "r",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.775,
+          "top5": 0.95,
+          "errors": 9
+        },
+        {
+          "label": "A",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.8,
+          "top5": 0.875,
+          "errors": 8
+        },
+        {
+          "label": "L",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.8,
+          "top5": 0.9,
+          "errors": 8
+        },
+        {
+          "label": "N",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.8,
+          "top5": 0.9,
+          "errors": 8
+        },
+        {
+          "label": "X",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.8,
+          "top5": 0.925,
+          "errors": 8
+        },
+        {
+          "label": "(",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.825,
+          "top5": 1.0,
+          "errors": 7
+        }
+      ],
+      "lowest_supported_labels": [
+        {
+          "label": ".",
+          "test_support": 21,
+          "training_support": 44,
+          "top1": 0.047619047619047616,
+          "top5": 0.2857142857142857,
+          "errors": 20
+        },
+        {
+          "label": "!",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.925,
+          "top5": 1.0,
+          "errors": 3
+        },
+        {
+          "label": "\"",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.3,
+          "top5": 0.925,
+          "errors": 28
+        },
+        {
+          "label": "%",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "'",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.05,
+          "top5": 0.675,
+          "errors": 38
+        },
+        {
+          "label": "(",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.825,
+          "top5": 1.0,
+          "errors": 7
+        },
+        {
+          "label": ")",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.925,
+          "top5": 1.0,
+          "errors": 3
+        },
+        {
+          "label": ",",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.5,
+          "top5": 0.65,
+          "errors": 20
+        },
+        {
+          "label": "-",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.925,
+          "top5": 1.0,
+          "errors": 3
+        },
+        {
+          "label": "0",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.725,
+          "top5": 1.0,
+          "errors": 11
+        },
+        {
+          "label": "1",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.775,
+          "top5": 1.0,
+          "errors": 9
+        },
+        {
+          "label": "2",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.95,
+          "top5": 1.0,
+          "errors": 2
+        },
+        {
+          "label": "3",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": "4",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.775,
+          "top5": 0.975,
+          "errors": 9
+        },
+        {
+          "label": "5",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.925,
+          "top5": 0.975,
+          "errors": 3
+        },
+        {
+          "label": "6",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.925,
+          "top5": 1.0,
+          "errors": 3
+        },
+        {
+          "label": "7",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.85,
+          "top5": 0.925,
+          "errors": 6
+        },
+        {
+          "label": "8",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.95,
+          "top5": 1.0,
+          "errors": 2
+        },
+        {
+          "label": "9",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.675,
+          "top5": 0.975,
+          "errors": 13
+        },
+        {
+          "label": ":",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.825,
+          "top5": 1.0,
+          "errors": 7
+        },
+        {
+          "label": ";",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.825,
+          "top5": 0.975,
+          "errors": 7
+        },
+        {
+          "label": "<",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 1.0,
+          "top5": 1.0,
+          "errors": 0
+        },
+        {
+          "label": ">",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.9,
+          "top5": 1.0,
+          "errors": 4
+        },
+        {
+          "label": "A",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.8,
+          "top5": 0.875,
+          "errors": 8
+        },
+        {
+          "label": "B",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.95,
+          "top5": 1.0,
+          "errors": 2
+        },
+        {
+          "label": "C",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.925,
+          "top5": 1.0,
+          "errors": 3
+        },
+        {
+          "label": "D",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.85,
+          "top5": 1.0,
+          "errors": 6
+        },
+        {
+          "label": "E",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.85,
+          "top5": 0.9,
+          "errors": 6
+        },
+        {
+          "label": "F",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.85,
+          "top5": 0.95,
+          "errors": 6
+        },
+        {
+          "label": "G",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.925,
+          "top5": 0.925,
+          "errors": 3
+        },
+        {
+          "label": "H",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.85,
+          "top5": 0.875,
+          "errors": 6
+        },
+        {
+          "label": "I",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.65,
+          "top5": 0.75,
+          "errors": 14
+        },
+        {
+          "label": "J",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.775,
+          "top5": 0.975,
+          "errors": 9
+        },
+        {
+          "label": "K",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.7,
+          "top5": 0.9,
+          "errors": 12
+        },
+        {
+          "label": "L",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.8,
+          "top5": 0.9,
+          "errors": 8
+        },
+        {
+          "label": "M",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.725,
+          "top5": 0.975,
+          "errors": 11
+        },
+        {
+          "label": "N",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.8,
+          "top5": 0.9,
+          "errors": 8
+        },
+        {
+          "label": "O",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.575,
+          "top5": 0.925,
+          "errors": 17
+        },
+        {
+          "label": "P",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.5,
+          "top5": 0.925,
+          "errors": 20
+        },
+        {
+          "label": "Q",
+          "test_support": 40,
+          "training_support": 72,
+          "top1": 0.95,
+          "top5": 0.95,
+          "errors": 2
+        }
+      ]
+    }
+  },
+  "case_size_proxy": {
+    "scope": "same-writer isolated-glyph relative-size proxy; not continuous-formula product evidence",
+    "writers": 49,
+    "by_source": {
+      "hwrt": {
+        "samples": 84,
+        "raw_correct": 45,
+        "resolved_correct": 33,
+        "changed": 18,
+        "beneficial": 3,
+        "harmful": 15,
+        "raw_top1": 0.5357142857142857,
+        "resolved_top1": 0.39285714285714285,
+        "delta_percentage_points": -14.285714285714286
+      },
+      "uci-uji-pen-v2": {
+        "samples": 640,
+        "raw_correct": 327,
+        "resolved_correct": 292,
+        "changed": 120,
+        "beneficial": 41,
+        "harmful": 76,
+        "raw_top1": 0.5109375,
+        "resolved_top1": 0.45625,
+        "delta_percentage_points": -5.46875
+      }
+    },
+    "product_validation": false
+  },
+  "checkpoint": "research\\runs\\math_ink_06_federation_clean_seed17_stage3_fullsource_b32_20260725\\math_ink_06_candidate.pt",
+  "device": "cuda",
+  "training_source_ids": [
+    "hwrt",
+    "uci-pendigits",
+    "uci-uji-pen-v2"
+  ],
+  "product_validation": false
+}

scripts/audit_math_ink_06_federation_errors.py

@@ -18,6 +18,10 @@ if str(PROJECT_ROOT / "scripts") not in sys.path:
     sys.path.insert(0, str(PROJECT_ROOT / "scripts"))
 
 from audit_math_ink_06_online_errors import summarize_online_predictions06
+from math_grid_drawer.research.case_context06 import (
+    SIZE_DEPENDENT_CASE_BASES06,
+    apply_relative_case_context06,
+)
 from math_grid_drawer.research.ink06_federation import (
     FederatedPairedInk06Dataset,
     load_product_federation06,
@@ -26,6 +30,86 @@ from math_grid_drawer.research.ink06_federation import (
 from math_grid_drawer.research.math_ink_06 import MathInk06Engine
 
 
+def _bbox_context06(record: dict) -> dict:
+    """필요 변수: 원본 source stroke·canvas. 작동 원리: 128 정규화 전에 있던 상대 높이를 context resolver용 bbox로 보존한다."""
+
+    points = []
+    for stroke in record.get("strokes") or []:
+        values = stroke.get("points", []) if isinstance(stroke, dict) else stroke
+        for point in values:
+            if isinstance(point, dict):
+                points.append((float(point["x"]), float(point["y"])))
+            else:
+                points.append((float(point[0]), float(point[1])))
+    if not points:
+        return dict(record)
+    canvas = record.get("canvas") or {}
+    height = max(float(canvas.get("height", 1.0)), 1e-6)
+    top = min(point[1] for point in points) / height
+    bottom = max(point[1] for point in points) / height
+    output = dict(record)
+    output["baseline_context"] = {
+        "bbox_top": top,
+        "bbox_bottom": bottom,
+        "bbox_height": max(bottom - top, 1e-6),
+    }
+    return output
+
+
+def _case_size_proxy06(
+    *, logits: torch.Tensor, labels: tuple[str, ...], records: list[dict],
+) -> dict:
+    """필요 변수: writer별 logits·원본 bbox. 작동 원리: 동일 writer를 행 anchor의 대리값으로 삼아 대소문자 상대크기 이득과 회귀를 함께 센다."""
+
+    enriched = [_bbox_context06(record) for record in records]
+    groups: dict[tuple[str, str], list[int]] = {}
+    for index, record in enumerate(enriched):
+        key = (str(record.get("source")), str(record.get("writer_key") or "missing"))
+        groups.setdefault(key, []).append(index)
+    predictions = logits.argmax(dim=1)
+    source_rows: dict[str, dict[str, int]] = {}
+    for (source, _writer), indices in groups.items():
+        group_logits = logits[indices]
+        group_records = [enriched[index] for index in indices]
+        resolved, decisions = apply_relative_case_context06(group_logits, labels, group_records)
+        raw = group_logits.argmax(dim=1)
+        revised = resolved.argmax(dim=1)
+        for local_index, global_index in enumerate(indices):
+            truth = str(records[global_index]["label"])
+            if len(truth) != 1 or not truth.isascii() or truth.lower() not in SIZE_DEPENDENT_CASE_BASES06:
+                continue
+            row = source_rows.setdefault(source, {
+                "samples": 0, "raw_correct": 0, "resolved_correct": 0,
+                "changed": 0, "beneficial": 0, "harmful": 0,
+            })
+            raw_label = labels[int(raw[local_index])]
+            resolved_label = labels[int(revised[local_index])]
+            row["samples"] += 1
+            row["raw_correct"] += int(raw_label == truth)
+            row["resolved_correct"] += int(resolved_label == truth)
+            row["changed"] += int(raw_label != resolved_label)
+            row["beneficial"] += int(raw_label != truth and resolved_label == truth)
+            row["harmful"] += int(raw_label == truth and resolved_label != truth)
+    by_source = {
+        source: {
+            **row,
+            "raw_top1": row["raw_correct"] / row["samples"] if row["samples"] else 0.0,
+            "resolved_top1": row["resolved_correct"] / row["samples"] if row["samples"] else 0.0,
+            "delta_percentage_points": (
+                (row["resolved_correct"] - row["raw_correct"]) * 100.0 / row["samples"]
+                if row["samples"] else 0.0
+            ),
+        }
+        for source, row in sorted(source_rows.items())
+    }
+    return {
+        "scope": "same-writer isolated-glyph relative-size proxy; not continuous-formula product evidence",
+        "writers": len(groups),
+        "by_source": by_source,
+        "product_validation": False,
+    }
+
+
 def _evaluate_online06(
     engine: MathInk06Engine,
     records: list[dict],
@@ -134,6 +218,11 @@ def main() -> None:
             )
         }
     report["source_bottlenecks"] = source_bottlenecks
+    report["case_size_proxy"] = _case_size_proxy06(
+        logits=log_probabilities,
+        labels=engine.labels,
+        records=metadata,
+    )
     report.update({
         "checkpoint": str(args.checkpoint),
         "device": device,