Add reproducible LiteRT Colab conversion bundle

MANIFEST.json

@@ -1,11 +1,11 @@
 {
-    "schema":  "aiflow-hf-research-snapshot-v8",
-    "generated_at":  "2026-07-23T20:04:13.5124140Z",
+    "schema":  "aiflow-hf-research-snapshot-v9",
+    "generated_at":  "2026-07-23T20:15:19.8122156Z",
     "track":  "R_noncommercial_plus_rejected_P_proxy",
     "product_validation":  false,
     "public_release":  true,
     "contains_raw_dataset":  false,
-    "tests":  "282 passed",
+    "tests":  "284 passed",
     "retracted_paths":  [
                             "models/auxiliary/boundary_auxiliary_head.pt",
                             "models/auxiliary/seed17/boundary_joint_delta.pt",
@@ -30,28 +30,28 @@
                   },
                   {
                       "path":  "reports/RESEARCH_REPORT.md",
-                      "bytes":  33159,
-                      "sha256":  "19c06349b4316cb538a48fbe1cf5fccaaa386ca215615ae6ca53bfaefc0995df"
-                  },
-                  {
-                      "path":  "scripts/evaluate_math_ink_06_p_boundary_device_stress.py",
-                      "bytes":  10444,
-                      "sha256":  "bdd75047c4887a6e53720f1c84d224545844cdbd9b0239826a605d578b3c92a6"
+                      "bytes":  34116,
+                      "sha256":  "c2c1f9fc166525912d34f421b80b3f2662b18ec4a4d47934990c37b9432223e8"
                   },
                   {
                       "path":  "scripts/evaluate_crohme_tray_joint_selector.py",
                       "bytes":  11671,
                       "sha256":  "58c175a95cdeb9fc764fb96594a76888122b8f118f2f26da1283e743b1a7cd04"
                   },
+                  {
+                      "path":  "scripts/build_math_ink_06_litert_colab_bundle.py",
+                      "bytes":  6862,
+                      "sha256":  "f008635ab59344005f65f1394f56275e4c92610cd25781a2d95243d060010718"
+                  },
                   {
                       "path":  "scripts/audit_math_ink_06_local_baseline_overmerge.py",
                       "bytes":  11900,
                       "sha256":  "f32a18c8e7d3a0ea36379293e6be88bfb509b7df99846c38149fb8458e6cc994"
                   },
                   {
-                      "path":  "reports/online_casecontext_seed31.json",
-                      "bytes":  33576,
-                      "sha256":  "30e6ca83aa4c04d1d257b0221fea7ac726f8e09e91906fdecbc12a2bd4d4d34b"
+                      "path":  "reports/p_boundary_joint_seed17.json",
+                      "bytes":  6920,
+                      "sha256":  "7af3bb995041e5ea621d3c9535f3552e264860f12a7e6145637849f5cd0220a8"
                   },
                   {
                       "path":  "reports/online_casecontext_seed17.json",
@@ -64,14 +64,14 @@
                       "sha256":  "27b31da71b9f39f7ded38aa1ce3694911af140d22854543d4ae2b5f9e709b479"
                   },
                   {
-                      "path":  "reports/online_casecontext_seed47.json",
-                      "bytes":  33727,
-                      "sha256":  "c3dfd328d9ff0a55eed2cb74657ad87e5e164968fddd7fae8c8fd92e137a4557"
+                      "path":  "reports/local_baseline_overmerge_audit.json",
+                      "bytes":  1237409,
+                      "sha256":  "33a9ccafebd559ca8b173f011e9a0a6f495d37b8fdb2693673663479d3457071"
                   },
                   {
-                      "path":  "reports/p_boundary_joint_seed17.json",
-                      "bytes":  6920,
-                      "sha256":  "7af3bb995041e5ea621d3c9535f3552e264860f12a7e6145637849f5cd0220a8"
+                      "path":  "reports/online_casecontext_seed31.json",
+                      "bytes":  33576,
+                      "sha256":  "30e6ca83aa4c04d1d257b0221fea7ac726f8e09e91906fdecbc12a2bd4d4d34b"
                   },
                   {
                       "path":  "reports/p_boundary_device_stress_sharedfix_3seed.json",
@@ -83,6 +83,11 @@
                       "bytes":  16685,
                       "sha256":  "9d250e5c75d4f104971551c07a702810bf94fb5b6e5e587035a624933766356e"
                   },
+                  {
+                      "path":  "reports/online_casecontext_seed47.json",
+                      "bytes":  33727,
+                      "sha256":  "c3dfd328d9ff0a55eed2cb74657ad87e5e164968fddd7fae8c8fd92e137a4557"
+                  },
                   {
                       "path":  "src/math_ink_06.py",
                       "bytes":  54708,
@@ -119,9 +124,9 @@
                       "sha256":  "8cc2dc1e13b56c21d53a76b9b863cdf5c8f8dd315da6eb18c1bd2b5b87c1a616"
                   },
                   {
-                      "path":  "scripts/summarize_math_ink_06_p_boundary_joint.py",
-                      "bytes":  4580,
-                      "sha256":  "c6ce78f521232bdb7e967b0c358232a72dd3f79747f9e32eb240d9bb10c6a422"
+                      "path":  "scripts/train_math_ink_06_p_boundary_auxiliary.py",
+                      "bytes":  20688,
+                      "sha256":  "70ca8a05dcbd48e79d812cd7ff44ee3d8479055f898143daa6d6ea7580b2973e"
                   },
                   {
                       "path":  "scripts/preflight_math_ink_06_p_formula.py",
@@ -130,18 +135,18 @@
                   },
                   {
                       "path":  "scripts/export_math_ink_06_litert.py",
-                      "bytes":  8637,
-                      "sha256":  "e96b30ded57d33d4505b065c1ac69c5ac7db43618c0eb7f3b7fb787b6fcaf1f6"
+                      "bytes":  11876,
+                      "sha256":  "3d5362341c29ef31493969ccf833c21cb1d4e02d8ccf9ccd217b038091f9e43c"
                   },
                   {
-                      "path":  "scripts/sweep_math_ink_06_component_competition_guard.py",
-                      "bytes":  9084,
-                      "sha256":  "e2dee616e2c6ebdc32281624889017685648b7076515242e04a1452d98d1c7da"
+                      "path":  "scripts/evaluate_math_ink_06_p_boundary_device_stress.py",
+                      "bytes":  10444,
+                      "sha256":  "bdd75047c4887a6e53720f1c84d224545844cdbd9b0239826a605d578b3c92a6"
                   },
                   {
-                      "path":  "scripts/train_math_ink_06_p_boundary_auxiliary.py",
-                      "bytes":  20688,
-                      "sha256":  "70ca8a05dcbd48e79d812cd7ff44ee3d8479055f898143daa6d6ea7580b2973e"
+                      "path":  "scripts/summarize_math_ink_06_p_boundary_joint.py",
+                      "bytes":  4580,
+                      "sha256":  "c6ce78f521232bdb7e967b0c358232a72dd3f79747f9e32eb240d9bb10c6a422"
                   },
                   {
                       "path":  "scripts/train_math_ink_06_boundary_behavior_guard.py",
@@ -154,14 +159,9 @@
                       "sha256":  "28cfc7b9b695256cfa9199d912d1d938d402292b88e9411d42fddb3508446d3b"
                   },
                   {
-                      "path":  "exports/seed17/online.pt2",
-                      "bytes":  9192090,
-                      "sha256":  "d8a4fa6a631d5f1f4e8871b5c551d647530312260caace3cab4bca01004f7891"
-                  },
-                  {
-                      "path":  "exports/seed17/export_manifest.json",
-                      "bytes":  1030,
-                      "sha256":  "48a224d03fb2e8f1c4fcb6ca3f61456521f4437b4cfc845572111cd84cb93833"
+                      "path":  "scripts/sweep_math_ink_06_component_competition_guard.py",
+                      "bytes":  9084,
+                      "sha256":  "e2dee616e2c6ebdc32281624889017685648b7076515242e04a1452d98d1c7da"
                   },
                   {
                       "path":  "contracts/aiflow_p_formula_v1.schema.json",
@@ -169,25 +169,40 @@
                       "sha256":  "c78c8a612d2f24a5c09fe1aa158b365481102869c3f84b3a5510a895b416c424"
                   },
                   {
-                      "path":  "exports/seed17/raster.pt2",
-                      "bytes":  9465689,
-                      "sha256":  "419301d67bbfe7b807e1ea676a33827c749f632300392d54eaf48c701ca07512"
+                      "path":  "colab/aiflow_math_ink_06_litert_bundle.zip",
+                      "bytes":  17276380,
+                      "sha256":  "e20ed37a7c35daabaae72b6108204ae08da3b909a19f230528bb40ab198c7ca8"
                   },
                   {
-                      "path":  "models/seed17/online_adapter.pt",
-                      "bytes":  2875654,
-                      "sha256":  "8ea224d7bc107a4c8b6cb91c9de6598000dae463d3e6d39480c081ea728cc87f"
+                      "path":  "colab/aiflow_math_ink_06_litert_bundle.manifest.json",
+                      "bytes":  2978,
+                      "sha256":  "03a7ebfe56cc8855494ecfafb463290d4bbe9b0fb0582153e17b62ea6726bdd6"
                   },
                   {
-                      "path":  "models/seed17/behavior_role_head.pt",
-                      "bytes":  75794,
-                      "sha256":  "57ced82a2600bd342bba68ba0636ecc3fc98c3dc461e26a132b5264d0a3d95bd"
+                      "path":  "exports/seed17/export_manifest.json",
+                      "bytes":  1031,
+                      "sha256":  "06b79357ca9c80934925d8e7d59a81c3a9c1f4fdc017ca4a6ed19ff6580ed6ab"
                   },
                   {
                       "path":  "models/seed17/base_378.pt",
                       "bytes":  8638290,
                       "sha256":  "f3b0e53b50e41cf57d4881c501371b35648cf049d603984a0edf6148c2b62bd7"
                   },
+                  {
+                      "path":  "exports/seed17/raster.pt2",
+                      "bytes":  14380889,
+                      "sha256":  "f2b3f0612a7016f44bf9271399fd2ab522048bc301e75dbeb52c21b31356dfbd"
+                  },
+                  {
+                      "path":  "exports/seed17/online.pt2",
+                      "bytes":  9921690,
+                      "sha256":  "fb07c249e49720310a0c627adfc1c21fa75f7203332e20b202622cbe05fa0214"
+                  },
+                  {
+                      "path":  "colab/AIFlow_Math_Ink_06_LiteRT.ipynb",
+                      "bytes":  3505,
+                      "sha256":  "7c0a2118bc84576957048e882df33681eb9a4e2fe832011b4d73117c0b57b0c1"
+                  },
                   {
                       "path":  "pyproject.snapshot.toml",
                       "bytes":  669,
@@ -200,13 +215,13 @@
                   },
                   {
                       "path":  "MODEL_INDEX.json",
-                      "bytes":  2406,
-                      "sha256":  "2419245ba889772d562d524999e9d14f801e70d87ce75e38231722eb5aecd697"
+                      "bytes":  2473,
+                      "sha256":  "8926674c956664b98d2fc1519e75355ad13efe57228d122ce208defebdfb90f1"
                   },
                   {
                       "path":  "README.md",
-                      "bytes":  11248,
-                      "sha256":  "d102514283972dbddc061848a1fb33a5d4f581dea0b43dc9eb3f4f8de46d5885"
+                      "bytes":  11411,
+                      "sha256":  "5e8ee6eef8985af8404540edc7fd000d5b03a095ed1eac05cabcfd8d346ab064"
                   },
                   {
                       "path":  "assets/stroke_encoding.svg",
@@ -223,11 +238,6 @@
                       "bytes":  824191,
                       "sha256":  "18aff472c1bdb25a519b1553364229c9869aed1f36a9d663cf88a1bc808fab8f"
                   },
-                  {
-                      "path":  "reports/behavior_role_seed47.json",
-                      "bytes":  30091,
-                      "sha256":  "3d4dcbe2f2b0afb483f992dacd62aab1089bdc41d55cc259bcb7fc21b5294db0"
-                  },
                   {
                       "path":  "reports/behavior_role_seed31.json",
                       "bytes":  45412,
@@ -239,14 +249,14 @@
                       "sha256":  "e2072fec1b1065777f0e9218c48d27f096b22f5545e7cc8d9962939282719561"
                   },
                   {
-                      "path":  "reports/boundary_behavior_guard_report.json",
-                      "bytes":  72053,
-                      "sha256":  "f55a230a9f0892b9676c71db46e9e527862b3d8da776dc43db767ce7a0f6a5c7"
+                      "path":  "reports/behavior_role_3seed_summary.json",
+                      "bytes":  3567,
+                      "sha256":  "b16049218d2d006d952f59f55642fa9f3fbfbc3cc3a730fbd7b075f507c3b251"
                   },
                   {
-                      "path":  "reports/local_baseline_overmerge_audit.json",
-                      "bytes":  1237409,
-                      "sha256":  "33a9ccafebd559ca8b173f011e9a0a6f495d37b8fdb2693673663479d3457071"
+                      "path":  "reports/behavior_role_seed47.json",
+                      "bytes":  30091,
+                      "sha256":  "3d4dcbe2f2b0afb483f992dacd62aab1089bdc41d55cc259bcb7fc21b5294db0"
                   },
                   {
                       "path":  "reports/local_baseline_guard_report.json",
@@ -259,14 +269,14 @@
                       "sha256":  "4d3db5f71e5008381dd072f2c10aafb22e08f96e23f2a32f82a3a46abd9c70b9"
                   },
                   {
-                      "path":  "models/seed31/online_adapter.pt",
-                      "bytes":  2875654,
-                      "sha256":  "fedbbab64cea6755e76e0633367ee45fbcc1d5726bb9dfdb7c5fe4bbc968bdf2"
+                      "path":  "reports/boundary_behavior_guard_report.json",
+                      "bytes":  72053,
+                      "sha256":  "f55a230a9f0892b9676c71db46e9e527862b3d8da776dc43db767ce7a0f6a5c7"
                   },
                   {
-                      "path":  "models/seed31/behavior_role_head.pt",
-                      "bytes":  75794,
-                      "sha256":  "b7c22cf49320b90a6fd103d39cc3ff08ea7734ab83275aad66e2b529b749d9ae"
+                      "path":  "models/seed47/online_adapter.pt",
+                      "bytes":  2876038,
+                      "sha256":  "46394eeebad3d27d3178f78e693e160700ba3f454d79044714b835bc51c07056"
                   },
                   {
                       "path":  "models/seed31/base_378.pt",
@@ -274,24 +284,34 @@
                       "sha256":  "e3c57bc8cbbd4542cacf7ab9200a9031dcdc5ee63ffbd6ad6c9f7ef4f0a64f37"
                   },
                   {
-                      "path":  "models/seed47/base_378.pt",
-                      "bytes":  8638290,
-                      "sha256":  "4131b21f6b00471c1c5a82aee4a2da45e599e42dc742fc1dd0dd8afb970f8939"
+                      "path":  "models/seed17/online_adapter.pt",
+                      "bytes":  2875654,
+                      "sha256":  "8ea224d7bc107a4c8b6cb91c9de6598000dae463d3e6d39480c081ea728cc87f"
                   },
                   {
-                      "path":  "reports/behavior_role_3seed_summary.json",
-                      "bytes":  3567,
-                      "sha256":  "b16049218d2d006d952f59f55642fa9f3fbfbc3cc3a730fbd7b075f507c3b251"
+                      "path":  "models/seed17/behavior_role_head.pt",
+                      "bytes":  75794,
+                      "sha256":  "57ced82a2600bd342bba68ba0636ecc3fc98c3dc461e26a132b5264d0a3d95bd"
                   },
                   {
-                      "path":  "models/seed47/online_adapter.pt",
-                      "bytes":  2876038,
-                      "sha256":  "46394eeebad3d27d3178f78e693e160700ba3f454d79044714b835bc51c07056"
+                      "path":  "models/seed31/behavior_role_head.pt",
+                      "bytes":  75794,
+                      "sha256":  "b7c22cf49320b90a6fd103d39cc3ff08ea7734ab83275aad66e2b529b749d9ae"
                   },
                   {
                       "path":  "models/seed47/behavior_role_head.pt",
                       "bytes":  75794,
                       "sha256":  "f28fad4e0289e3bc9400829f41aba9bbd75c43d445b7462a84aadbdbe635c10b"
+                  },
+                  {
+                      "path":  "models/seed47/base_378.pt",
+                      "bytes":  8638290,
+                      "sha256":  "4131b21f6b00471c1c5a82aee4a2da45e599e42dc742fc1dd0dd8afb970f8939"
+                  },
+                  {
+                      "path":  "models/seed31/online_adapter.pt",
+                      "bytes":  2875654,
+                      "sha256":  "fedbbab64cea6755e76e0633367ee45fbcc1d5726bb9dfdb7c5fe4bbc968bdf2"
                   }
               ]
 }

MODEL_INDEX.json

@@ -64,6 +64,8 @@
     "shared_state_applied": true,
     "top1_agreement": 1.0,
     "maximum_absolute_logit_error": 0.0,
+    "total_pt2_bytes": 24302579,
+    "representative_samples": 76,
     "litert_flatbuffer": false
   },
   "release_state": {

README.md

@@ -191,7 +191,7 @@ Joint delta 없는 main+shadow auxiliary device stress에서 최악 exact/family
 
 ### Composite torch.export
 
-Export 그래프는 이제 base-only가 아니라 online/raster modality adapter와 shared state를 포함한다. 세 seed 모두 각 경로 대표 입력 76개에서 eager 대비 top-1 100% 일치, 최대 logit 절대오차 0.0을 기록했다. Seed-17의 online/raster `.pt2` 합계는 18,657,779 bytes다.
+Export 그래프는 이제 base-only가 아니라 online/raster modality adapter와 shared state를 포함한다. 세 seed 모두 각 경로 대표 입력 76개에서 eager 대비 top-1 100% 일치, 최대 logit 절대오차 0.0을 기록했다. Stale output 재사용을 atomic replace로 수정한 fresh seed-17 online/raster `.pt2` 합계는 24,302,579 bytes다.
 
 `.pt2`는 Android용 `.tflite`가 아니다. LiteRT Torch 0.9.1 변환과 Android runtime parity는 아직 완료되지 않았으므로 `litert_exported=false`를 유지한다.
 
@@ -247,5 +247,6 @@ PyTorch checkpoint와 joblib/pickle은 신뢰할 수 없는 출처에서 로드
 - P boundary 정정 3-seed: [`reports/p_boundary_joint_sharedfix_3seed_summary.json`](reports/p_boundary_joint_sharedfix_3seed_summary.json)
 - 정정 main device stress: [`reports/p_boundary_device_stress_sharedfix_3seed.json`](reports/p_boundary_device_stress_sharedfix_3seed.json)
 - Seed-17 composite export: [`exports/seed17/export_manifest.json`](exports/seed17/export_manifest.json)
+- LiteRT Colab notebook: [`colab/AIFlow_Math_Ink_06_LiteRT.ipynb`](colab/AIFlow_Math_Ink_06_LiteRT.ipynb)
 - 실제 P formula schema: [`contracts/aiflow_p_formula_v1.schema.json`](contracts/aiflow_p_formula_v1.schema.json)
 - 파일 checksum: [`MANIFEST.json`](MANIFEST.json)

colab/AIFlow_Math_Ink_06_LiteRT.ipynb

@@ -0,0 +1,95 @@
+{
+  "nbformat": 4,
+  "nbformat_minor": 5,
+  "metadata": {
+    "colab": {"name": "AIFlow Math Ink 0.6 LiteRT"},
+    "kernelspec": {"display_name": "Python 3", "name": "python3"}
+  },
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "# AIFlow Math Ink 0.6 — LiteRT 변환\n",
+        "\n",
+        "공개 seed-17 composite 모델을 공식 LiteRT Torch 0.9.1로 변환하고 실제 대표 입력 76개의 PyTorch/LiteRT parity를 검사합니다. 출력 gate를 통과하기 전에는 Android 배포 artifact로 취급하지 않습니다."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "!pip -q install litert-torch==0.9.1 pillow huggingface-hub\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "from huggingface_hub import hf_hub_download\n",
+        "from pathlib import Path\n",
+        "from zipfile import ZipFile\n",
+        "import hashlib, json\n",
+        "\n",
+        "repo = 'cwLeeDev/aiflow-math-ink-06-intermediate'\n",
+        "archive = Path(hf_hub_download(repo_id=repo, filename='colab/aiflow_math_ink_06_litert_bundle.zip'))\n",
+        "root = Path('/content/aiflow_math_ink_06_litert')\n",
+        "root.mkdir(parents=True, exist_ok=True)\n",
+        "with ZipFile(archive) as bundle:\n",
+        "    bundle.extractall(root)\n",
+        "manifest = json.loads((root / 'LITERT_COLAB_BUNDLE_MANIFEST.json').read_text(encoding='utf-8'))\n",
+        "for row in manifest['files']:\n",
+        "    payload = (root / row['path']).read_bytes()\n",
+        "    assert len(payload) == row['bytes']\n",
+        "    assert hashlib.sha256(payload).hexdigest() == row['sha256']\n",
+        "print('bundle verified:', len(manifest['files']), 'files')\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "%cd /content/aiflow_math_ink_06_litert\n",
+        "!PYTHONPATH=src python scripts/export_math_ink_06_litert.py \\\n",
+        "  --checkpoint artifacts/base_378.pt \\\n",
+        "  --adapter-checkpoint artifacts/online_adapter.pt \\\n",
+        "  --representative-inputs artifacts/representative_inputs.pt \\\n",
+        "  --convert-litert --output outputs\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "report = json.loads((root / 'outputs/export_manifest.json').read_text(encoding='utf-8'))\n",
+        "assert report['torch_export_gate_passed']\n",
+        "assert report['litert']['online']['gate_passed']\n",
+        "assert report['litert']['raster']['gate_passed']\n",
+        "total_bytes = report['litert']['online']['bytes'] + report['litert']['raster']['bytes']\n",
+        "assert total_bytes <= 25 * 1024 * 1024\n",
+        "print('LiteRT gate passed, total bytes:', total_bytes)\n",
+        "report\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "from google.colab import files\n",
+        "import shutil\n",
+        "shutil.make_archive('/content/aiflow_math_ink_06_litert_outputs', 'zip', root / 'outputs')\n",
+        "files.download('/content/aiflow_math_ink_06_litert_outputs.zip')\n"
+      ]
+    }
+  ]
+}

colab/aiflow_math_ink_06_litert_bundle.manifest.json

@@ -0,0 +1,85 @@
+{
+  "schema": "aiflow-math-ink-06-litert-colab-bundle-v1",
+  "generated_at": "2026-07-23T20:13:15.384782+00:00",
+  "seed": 17,
+  "litert_torch_version": "0.9.1",
+  "representative_samples": 76,
+  "files": [
+    {
+      "path": "pyproject.toml",
+      "bytes": 669,
+      "sha256": "e9571da506b96d9ab6baa8f8e2c1d87cd8ee9fd0c880f3cf6646a99086ae3f09"
+    },
+    {
+      "path": "src/math_grid_drawer/__init__.py",
+      "bytes": 96,
+      "sha256": "16625f83572539172333f4146ceaf255f242b879010a9a2a44bc85cfb651396a"
+    },
+    {
+      "path": "src/math_grid_drawer/research/__init__.py",
+      "bytes": 91,
+      "sha256": "51bf186de1b9da930191d5d25d172c7e3372204784326c942b38f7af0775edf0"
+    },
+    {
+      "path": "src/math_grid_drawer/research/ink06_canonical.py",
+      "bytes": 14198,
+      "sha256": "3bd7337f222eec7355e5bdc6d485dc654549978aebbd5e0de87c19222373a17c"
+    },
+    {
+      "path": "src/math_grid_drawer/research/ink06_export.py",
+      "bytes": 4474,
+      "sha256": "37c67827f1c39df9cb0f6c45f782c58a2d1bdf5190e51070de817d1ba2693b67"
+    },
+    {
+      "path": "src/math_grid_drawer/research/math_ink_06.py",
+      "bytes": 54708,
+      "sha256": "6c41a7ff9958fe6bd396b3321586d7a935ef0206553ce633327962c2d27ffff4"
+    },
+    {
+      "path": "src/math_grid_drawer/research/raster_skeleton06.py",
+      "bytes": 6052,
+      "sha256": "d6807fc96c6e75e6bf12054382d29a227b1aa05fc867ab44badb6cca3fcb5cf2"
+    },
+    {
+      "path": "src/math_grid_drawer/research/skeleton_adapter06.py",
+      "bytes": 11697,
+      "sha256": "068cc8a616da25425827bda3aae4ae5d7a7619ec9623df77c6477447e5cf3d50"
+    },
+    {
+      "path": "src/math_grid_drawer/research/trajectory_sequence.py",
+      "bytes": 25290,
+      "sha256": "a3ad8193b6da8e9599c4291618983a40808a264075d2ecb3f205c711c3c5d032"
+    },
+    {
+      "path": "scripts/export_math_ink_06_litert.py",
+      "bytes": 11876,
+      "sha256": "3d5362341c29ef31493969ccf833c21cb1d4e02d8ccf9ccd217b038091f9e43c"
+    },
+    {
+      "path": "artifacts/base_378.pt",
+      "bytes": 8638290,
+      "sha256": "f3b0e53b50e41cf57d4881c501371b35648cf049d603984a0edf6148c2b62bd7"
+    },
+    {
+      "path": "artifacts/online_adapter.pt",
+      "bytes": 2875654,
+      "sha256": "8ea224d7bc107a4c8b6cb91c9de6598000dae463d3e6d39480c081ea728cc87f"
+    },
+    {
+      "path": "artifacts/representative_inputs.pt",
+      "bytes": 5721772,
+      "sha256": "8eda6e57a6fc1d32aa16b579b847bdd864538293bf37d8c1e226de7e5a21a2b6"
+    }
+  ],
+  "product_validation": false,
+  "bundle": "research\\runs\\math_ink_06_litert_colab_20260724\\aiflow_math_ink_06_litert_bundle.zip",
+  "bundle_bytes": 17276380,
+  "bundle_sha256": "e20ed37a7c35daabaae72b6108204ae08da3b909a19f230528bb40ab198c7ca8",
+  "verification": {
+    "schema": "aiflow-math-ink-06-litert-colab-bundle-v1",
+    "files": 13,
+    "failures": [],
+    "passed": true,
+    "product_validation": false
+  }
+}

colab/aiflow_math_ink_06_litert_bundle.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e20ed37a7c35daabaae72b6108204ae08da3b909a19f230528bb40ab198c7ca8
+size 17276380

exports/seed17/export_manifest.json

@@ -12,7 +12,7 @@
       "max_absolute_logit_error": 0.0,
       "gate_passed": true,
       "path": "online.pt2",
-      "bytes": 9192090
+      "bytes": 9921690
     },
     "raster": {
       "samples": 76,
@@ -21,7 +21,7 @@
       "max_absolute_logit_error": 0.0,
       "gate_passed": true,
       "path": "raster.pt2",
-      "bytes": 9465689
+      "bytes": 14380889
     }
   },
   "litert_package_available": false,

exports/seed17/online.pt2

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d8a4fa6a631d5f1f4e8871b5c551d647530312260caace3cab4bca01004f7891
-size 9192090
+oid sha256:fb07c249e49720310a0c627adfc1c21fa75f7203332e20b202622cbe05fa0214
+size 9921690

exports/seed17/raster.pt2

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:419301d67bbfe7b807e1ea676a33827c749f632300392d54eaf48c701ca07512
-size 9465689
+oid sha256:f2b3f0612a7016f44bf9271399fd2ab522048bc301e75dbeb52c21b31356dfbd
+size 14380889

reports/RESEARCH_REPORT.md

@@ -407,7 +407,9 @@ Joint delta 없이 정정된 main과 새 auxiliary head만 대상으로 device s
 
 기존 export script도 base checkpoint만 읽어 실제 main adapter를 우회했다. Online은 `online adapter → shared trajectory encoder`, raster는 `virtual top-4 → raster adapter → shared trajectory encoder`가 그래프 안에 포함되도록 수정했다. Dual adapter는 입력 데이터에 따른 런타임 분기 대신 online/raster branch를 export 시점에 고정한다.
 
-Seed 17·31·47 모두 online/raster 각각 대표 입력 76개에서 eager 대비 top-1 100% 일치, 최대 logit 절대오차 0.0으로 strict `torch.export` gate를 통과했다. Seed당 두 graph 합계는 18,657,779 bytes로 25MB 예산 안이다. 다만 `.pt2`는 Android LiteRT artifact가 아니다. 공식 LiteRT Torch 0.9.1은 Linux 환경을 요구하며 WSL 변환 환경 설치 중 WSL filesystem I/O 오류가 발생해 `.tflite` 변환과 runtime parity는 아직 미완료다.
+Seed 17·31·47 모두 online/raster 각각 대표 입력 76개에서 eager 대비 top-1 100% 일치, 최대 logit 절대오차 0.0으로 strict `torch.export` gate를 통과했다. 기존 출력 경로를 재사용할 때 오래된 `.pt2`가 남는 문제를 발견해 `.part → atomic replace`로 수정했고, fresh seed-17 두 graph 합계는 24,302,579 bytes로 25MiB 예산 안이다. 다만 `.pt2`는 Android LiteRT artifact가 아니다. 공식 LiteRT Torch 0.9.1은 Linux 환경을 요구하며 WSL 변환 환경 설치 중 WSL filesystem I/O 오류가 발생해 `.tflite` 변환과 runtime parity는 아직 미완료다.
+
+Linux/Colab 변환을 위해 label이 없는 실제 HWRT-derived online/raster 대표 입력 76개를 5,721,772-byte cache로 고정했다. 최소 bundle은 composite base·adapter, 필요한 source 7개, converter와 대표 cache만 포함하며 13개 entry·17,276,380 bytes다. Bundle 내부 SHA-256 전체 검증과 별도 디렉터리 추출 후 composite export smoke를 통과했다. Colab notebook은 공식 `litert-torch==0.9.1` 설치, 공개 Hugging Face bundle 다운로드, 해시 검증, online/raster `.tflite` 변환, 76개 전체 top-1 100%·logit 오차≤0.02·합계≤25MiB gate를 자동 실행한다.
 
 ### 실제 P 연속식 데이터 계약
 
@@ -420,7 +422,7 @@ Seed 17·31·47 모두 online/raster 각각 대표 입력 76개에서 eager 대
 - 양수 canvas 크기
 - token과 원본 raw stroke가 포함된 정답 symbol group
 
-Origin·writer·device·source가 둘 이상의 split에 나타나면 제품 평가를 거부한다. Timestamp와 pressure가 없는 symbol은 삭제하거나 관측값으로 위장하지 않고 missing slice로 센다. 검증된 formula의 실제 symbol group은 boundary 음성, 인접한 두 symbol group 결합은 boundary 양성으로 만든다. 이 계약은 실제 성능값을 만들지는 않지만, 향후 P 입력이 CROHME 정답 group이나 합성 고립기호 proxy와 섞이는 것을 방지한다. 전체 회귀는 282개가 통과했다.
+Origin·writer·device·source가 둘 이상의 split에 나타나면 제품 평가를 거부한다. Timestamp와 pressure가 없는 symbol은 삭제하거나 관측값으로 위장하지 않고 missing slice로 센다. 검증된 formula의 실제 symbol group은 boundary 음성, 인접한 두 symbol group 결합은 boundary 양성으로 만든다. 이 계약은 실제 성능값을 만들지는 않지만, 향후 P 입력이 CROHME 정답 group이나 합성 고립기호 proxy와 섞이는 것을 방지한다. 전체 회귀는 284개가 통과했다.
 
 ## 산출물
 
@@ -445,12 +447,15 @@ Origin·writer·device·source가 둘 이상의 split에 나타나면 제품 평
 - `scripts/summarize_math_ink_06_p_boundary_joint.py`
 - `scripts/evaluate_math_ink_06_p_boundary_device_stress.py`
 - `scripts/export_math_ink_06_litert.py`
+- `scripts/build_math_ink_06_litert_colab_bundle.py`
 - `scripts/preflight_math_ink_06_p_formula.py`
 - `scripts/analyze_crohme_lattice_failures.py`
 - `tests/test_behavior_context06.py`
 - `tests/test_behavior_role_head06.py`
 - `tests/test_p_formula_gate06.py`
 - `research/contracts/aiflow_p_formula_v1.schema.json`
+- `research/colab/AIFlow_Math_Ink_06_LiteRT.ipynb`
+- `research/runs/math_ink_06_litert_colab_20260724/aiflow_math_ink_06_litert_bundle.zip`
 - `research/AIFlow-MATH-INK-0.6-BEHAVIOR-CONTEXT-REPORT-20260724.md`
 - `research/runs/math_ink_06_behavior_role_3seed_20260724/run_summary.json`
 - `research/runs/math_ink_06_behavior_grouping_audit_20260724/report.json`

scripts/build_math_ink_06_litert_colab_bundle.py

@@ -0,0 +1,145 @@
+"""AIFlow Math Ink 0.6 LiteRT 변환에 필요한 최소 Colab ZIP을 결정적으로 만든다."""
+
+from __future__ import annotations
+
+import argparse
+from datetime import datetime, timezone
+from hashlib import sha256
+import json
+from pathlib import Path, PurePosixPath
+from zipfile import ZIP_DEFLATED, ZIP_STORED, ZipFile, ZipInfo
+
+
+PROJECT_ROOT = Path(__file__).parents[1]
+MANIFEST_NAME_06 = "LITERT_COLAB_BUNDLE_MANIFEST.json"
+
+
+def _sha256_bytes06(value: bytes) -> str:
+    """필요 변수: file bytes. 작동 원리: bundle entry와 manifest 검증용 SHA-256을 반환한다."""
+
+    return sha256(value).hexdigest()
+
+
+def _bundle_files06(
+    base_checkpoint: Path, adapter_checkpoint: Path, representative_inputs: Path,
+) -> dict[str, Path]:
+    """필요 변수: seed-17 composite artifact. 작동 원리: converter가 import하는 닫힌 최소 파일 집합을 매핑한다."""
+
+    return {
+        "pyproject.toml": PROJECT_ROOT / "pyproject.toml",
+        "src/math_grid_drawer/__init__.py": PROJECT_ROOT / "research/colab/litert_bundle_package_init.py",
+        "src/math_grid_drawer/research/__init__.py": PROJECT_ROOT / "src/math_grid_drawer/research/__init__.py",
+        "src/math_grid_drawer/research/ink06_canonical.py": PROJECT_ROOT / "src/math_grid_drawer/research/ink06_canonical.py",
+        "src/math_grid_drawer/research/ink06_export.py": PROJECT_ROOT / "src/math_grid_drawer/research/ink06_export.py",
+        "src/math_grid_drawer/research/math_ink_06.py": PROJECT_ROOT / "src/math_grid_drawer/research/math_ink_06.py",
+        "src/math_grid_drawer/research/raster_skeleton06.py": PROJECT_ROOT / "src/math_grid_drawer/research/raster_skeleton06.py",
+        "src/math_grid_drawer/research/skeleton_adapter06.py": PROJECT_ROOT / "src/math_grid_drawer/research/skeleton_adapter06.py",
+        "src/math_grid_drawer/research/trajectory_sequence.py": PROJECT_ROOT / "src/math_grid_drawer/research/trajectory_sequence.py",
+        "scripts/export_math_ink_06_litert.py": PROJECT_ROOT / "scripts/export_math_ink_06_litert.py",
+        "artifacts/base_378.pt": base_checkpoint,
+        "artifacts/online_adapter.pt": adapter_checkpoint,
+        "artifacts/representative_inputs.pt": representative_inputs,
+    }
+
+
+def build_litert_colab_bundle06(
+    base_checkpoint: Path,
+    adapter_checkpoint: Path,
+    representative_inputs: Path,
+    output: Path,
+) -> dict:
+    """필요 변수: composite checkpoint·대표 입력·ZIP 출력. 작동 원리: 경로·크기·해시가 고정된 Linux 변환 bundle을 만든다."""
+
+    files = _bundle_files06(base_checkpoint, adapter_checkpoint, representative_inputs)
+    missing = [name for name, path in files.items() if not path.is_file()]
+    if missing:
+        raise FileNotFoundError(f"LiteRT bundle 필수 파일이 없습니다: {missing}")
+    entries = []
+    payloads: dict[str, bytes] = {}
+    for name, path in files.items():
+        payload = path.read_bytes()
+        payloads[name] = payload
+        entries.append({
+            "path": name, "bytes": len(payload), "sha256": _sha256_bytes06(payload),
+        })
+    manifest = {
+        "schema": "aiflow-math-ink-06-litert-colab-bundle-v1",
+        "generated_at": datetime.now(timezone.utc).isoformat(),
+        "seed": 17,
+        "litert_torch_version": "0.9.1",
+        "representative_samples": 76,
+        "files": entries,
+        "product_validation": False,
+    }
+    output.parent.mkdir(parents=True, exist_ok=True)
+    temporary = output.with_suffix(output.suffix + ".part")
+    with ZipFile(temporary, "w", allowZip64=True) as bundle:
+        for name, payload in payloads.items():
+            compression = ZIP_STORED if PurePosixPath(name).suffix in {".pt"} else ZIP_DEFLATED
+            info = ZipInfo(name, date_time=(1980, 1, 1, 0, 0, 0))
+            info.compress_type = compression
+            info.external_attr = 0o644 << 16
+            bundle.writestr(info, payload)
+        info = ZipInfo(MANIFEST_NAME_06, date_time=(1980, 1, 1, 0, 0, 0))
+        info.compress_type = ZIP_DEFLATED
+        info.external_attr = 0o644 << 16
+        bundle.writestr(
+            info, json.dumps(manifest, ensure_ascii=False, indent=2).encode("utf-8") + b"\n",
+        )
+    temporary.replace(output)
+    manifest["bundle"] = str(output)
+    manifest["bundle_bytes"] = output.stat().st_size
+    manifest["bundle_sha256"] = sha256(output.read_bytes()).hexdigest()
+    return manifest
+
+
+def verify_litert_colab_bundle06(bundle_path: Path) -> dict:
+    """필요 변수: 생성 ZIP. 작동 원리: 추출 없이 모든 entry의 안전 경로·크기·SHA-256을 재검증한다."""
+
+    failures = []
+    with ZipFile(bundle_path) as bundle:
+        manifest = json.loads(bundle.read(MANIFEST_NAME_06).decode("utf-8"))
+        names = set(bundle.namelist())
+        for row in manifest["files"]:
+            name = str(row["path"])
+            path = PurePosixPath(name)
+            if path.is_absolute() or ".." in path.parts:
+                failures.append({"path": name, "reason": "unsafe_path"})
+            elif name not in names:
+                failures.append({"path": name, "reason": "missing"})
+            else:
+                payload = bundle.read(name)
+                if len(payload) != int(row["bytes"]):
+                    failures.append({"path": name, "reason": "bytes"})
+                elif _sha256_bytes06(payload) != str(row["sha256"]):
+                    failures.append({"path": name, "reason": "sha256"})
+    return {
+        "schema": manifest["schema"], "files": len(manifest["files"]),
+        "failures": failures, "passed": not failures, "product_validation": False,
+    }
+
+
+def main() -> None:
+    """필요 변수: CLI artifact 경로. 작동 원리: bundle 생성 직후 archive 자체 검증까지 수행한다."""
+
+    parser = argparse.ArgumentParser(description="Build Math Ink 0.6 LiteRT Colab bundle")
+    parser.add_argument("--base-checkpoint", type=Path, required=True)
+    parser.add_argument("--adapter-checkpoint", type=Path, required=True)
+    parser.add_argument("--representative-inputs", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    report = build_litert_colab_bundle06(
+        args.base_checkpoint, args.adapter_checkpoint, args.representative_inputs, args.output,
+    )
+    report["verification"] = verify_litert_colab_bundle06(args.output)
+    if not report["verification"]["passed"]:
+        raise ValueError(f"LiteRT Colab bundle 검증 실패: {report['verification']['failures']}")
+    report_path = args.output.with_suffix(args.output.suffix + ".manifest.json")
+    report_path.write_text(
+        json.dumps(report, ensure_ascii=False, indent=2) + "\n", encoding="utf-8",
+    )
+    print(json.dumps(report, ensure_ascii=False, indent=2))
+
+
+if __name__ == "__main__":
+    main()

scripts/export_math_ink_06_litert.py

@@ -16,7 +16,6 @@ SOURCE_ROOT = PROJECT_ROOT / "src"
 if str(SOURCE_ROOT) not in sys.path:
     sys.path.insert(0, str(SOURCE_ROOT))
 
-from math_grid_drawer.research.external_corpus import read_jsonl
 from math_grid_drawer.research.ink06_canonical import canonicalize_ink06, render_canonical_ink
 from math_grid_drawer.research.ink06_export import (
     OnlineExportWrapper06, RasterExportWrapper06, exported_equivalence06,
@@ -31,6 +30,8 @@ from math_grid_drawer.research.math_ink_06 import VirtualTrajectoryAdapter06
 def _representative_inputs(baseline_report: Path, data_path: Path) -> tuple[list[tuple[torch.Tensor, ...]], list[tuple[torch.Tensor, ...]]]:
     """필요 변수: strict baseline·HWRT JSONL. 작동 원리: 고정 76개를 128×19와 128×128 대표 입력으로 재구성한다."""
 
+    from math_grid_drawer.research.external_corpus import read_jsonl
+
     baseline = json.loads(baseline_report.read_text(encoding="utf-8"))
     accepted_ids = {row["sample_id"] for row in baseline["rows"] if row["raster_gate"]}
     records = {row["sample_id"]: row for row in read_jsonl(data_path) if row["sample_id"] in accepted_ids}
@@ -47,26 +48,77 @@ def _representative_inputs(baseline_report: Path, data_path: Path) -> tuple[list
     return online_inputs, raster_inputs
 
 
-def _convert_litert(wrapper: torch.nn.Module, sample: tuple[torch.Tensor, ...], output: Path) -> dict:
-    """필요 변수: export 호환 wrapper·표본·출력. 작동 원리: 공식 litert_torch.convert API로 flatbuffer를 만들고 1표본 오차를 검사한다."""
+def _load_representative_inputs06(
+    cache_path: Path,
+) -> tuple[list[tuple[torch.Tensor, ...]], list[tuple[torch.Tensor, ...]]]:
+    """필요 변수: 고정 representative cache. 작동 원리: Colab에서도 같은 76개 batch-1 입력을 복원한다."""
+
+    payload = torch.load(cache_path, map_location="cpu", weights_only=True)
+    if payload.get("schema") != "aiflow-math-ink-06-export-inputs-v1":
+        raise ValueError("지원하지 않는 export representative cache입니다.")
+    online, raster = payload["online"], payload["raster"]
+    if online.ndim != 3 or online.shape[1:] != (128, 19):
+        raise ValueError(f"online representative shape가 다릅니다: {tuple(online.shape)}")
+    if raster.ndim != 4 or raster.shape[1:] != (1, 128, 128):
+        raise ValueError(f"raster representative shape가 다릅니다: {tuple(raster.shape)}")
+    if len(online) != len(raster) or not len(online):
+        raise ValueError("online/raster representative 분모가 다릅니다.")
+    return (
+        [(value.unsqueeze(0),) for value in online],
+        [(value.unsqueeze(0),) for value in raster],
+    )
+
+
+def _save_representative_inputs06(
+    output: Path,
+    online_inputs: list[tuple[torch.Tensor, ...]],
+    raster_inputs: list[tuple[torch.Tensor, ...]],
+) -> None:
+    """필요 변수: 두 대표 입력 목록·출력. 작동 원리: 실제 HWRT-derived 입력만 tensor cache로 고정한다."""
+
+    torch.save({
+        "schema": "aiflow-math-ink-06-export-inputs-v1",
+        "online": torch.cat([row[0] for row in online_inputs], dim=0).cpu(),
+        "raster": torch.cat([row[0] for row in raster_inputs], dim=0).cpu(),
+        "samples": len(online_inputs),
+        "contains_labels": False,
+        "product_validation": False,
+    }, output)
+
+
+def _convert_litert(
+    wrapper: torch.nn.Module, samples: list[tuple[torch.Tensor, ...]], output: Path,
+) -> dict:
+    """필요 변수: export 호환 wrapper·대표 입력 전체·출력. 작동 원리: 공식 converter 뒤 76개 top-1/logit parity를 검사한다."""
 
     import litert_torch  # type: ignore[import-not-found]
 
-    edge_model = litert_torch.convert(wrapper.eval(), sample)
-    eager = wrapper(*sample)
-    edge = edge_model(*sample)
-    eager_values = eager if isinstance(eager, tuple) else (eager,)
-    edge_values = edge if isinstance(edge, tuple) else (edge,)
-    max_error = max(
-        float(np.max(np.abs(left.detach().cpu().numpy() - np.asarray(right))))
-        for left, right in zip(eager_values, edge_values)
-    )
-    top1_equal = int(np.argmax(eager_values[0].detach().cpu().numpy())) == int(np.argmax(np.asarray(edge_values[0])))
+    edge_model = litert_torch.convert(wrapper.eval(), samples[0])
+    top1_matches = 0
+    max_error = 0.0
+    with torch.inference_mode():
+        for sample in samples:
+            eager = wrapper(*sample)
+            edge = edge_model(*sample)
+            eager_values = eager if isinstance(eager, tuple) else (eager,)
+            edge_values = edge if isinstance(edge, tuple) else (edge,)
+            if len(eager_values) != len(edge_values):
+                raise ValueError("PyTorch/LiteRT 출력 개수가 다릅니다.")
+            max_error = max(max_error, max(
+                float(np.max(np.abs(left.detach().cpu().numpy() - np.asarray(right))))
+                for left, right in zip(eager_values, edge_values)
+            ))
+            top1_matches += int(
+                np.argmax(eager_values[0].detach().cpu().numpy(), axis=-1)[0]
+                == np.argmax(np.asarray(edge_values[0]), axis=-1)[0]
+            )
     edge_model.export(str(output))
     return {
         "converted": True, "path": output.name, "bytes": output.stat().st_size,
-        "sample_top1_agreement": top1_equal, "sample_max_absolute_logit_error": max_error,
-        "sample_gate_passed": top1_equal and max_error <= 0.02,
+        "samples": len(samples), "top1_matches": top1_matches,
+        "top1_agreement": top1_matches / max(len(samples), 1),
+        "max_absolute_logit_error": max_error,
+        "gate_passed": top1_matches == len(samples) and max_error <= 0.02,
     }
 
 
@@ -103,12 +155,24 @@ def _adapter_branches06(adapter: torch.nn.Module) -> tuple[torch.nn.Module, torc
     return adapter, adapter
 
 
+def _save_exported_program06(exported: torch.export.ExportedProgram, output: Path) -> None:
+    """필요 변수: export program·목표 파일. 작동 원리: stale ZIP 재사용 없이 임시 파일을 원자적으로 교체한다."""
+
+    temporary = output.with_suffix(output.suffix + ".part")
+    if temporary.exists():
+        temporary.unlink()
+    torch.export.save(exported, temporary)
+    temporary.replace(output)
+
+
 def main() -> None:
     """필요 변수: checkpoint·strict 입력·출력·변환 선택. 작동 원리: 두 경로의 export/동등성/선택적 LiteRT 결과를 manifest로 고정한다."""
 
     parser = argparse.ArgumentParser(description="Export Math Ink 0.6 for LiteRT")
     parser.add_argument("--checkpoint", type=Path, required=True)
     parser.add_argument("--adapter-checkpoint", type=Path, required=True)
+    parser.add_argument("--representative-inputs", type=Path)
+    parser.add_argument("--save-representative-inputs", type=Path)
     parser.add_argument("--baseline-report", type=Path, default=PROJECT_ROOT / "research/runs/full_model_stage2_20260722/isolated_checkpoint.json")
     parser.add_argument("--data", type=Path, default=PROJECT_ROOT / "research/data/open_pretrain/hwrt_expanded_v2/hwrt_expanded.jsonl.gz")
     parser.add_argument("--output", type=Path, required=True)
@@ -124,13 +188,21 @@ def main() -> None:
         engine.model, adapter=raster_adapter,
         fusion_mode=str(fusion["mode"]), score_weight=float(fusion["score_weight"]),
     ).eval()
-    online_inputs, raster_inputs = _representative_inputs(args.baseline_report, args.data)
+    if args.representative_inputs:
+        online_inputs, raster_inputs = _load_representative_inputs06(args.representative_inputs)
+    else:
+        online_inputs, raster_inputs = _representative_inputs(args.baseline_report, args.data)
+    if args.save_representative_inputs:
+        args.save_representative_inputs.parent.mkdir(parents=True, exist_ok=True)
+        _save_representative_inputs06(
+            args.save_representative_inputs, online_inputs, raster_inputs,
+        )
     online_export = torch.export.export(online, online_inputs[0], strict=True)
     raster_export = torch.export.export(raster, raster_inputs[0], strict=True)
     args.output.mkdir(parents=True, exist_ok=True)
     online_path, raster_path = args.output / "online.pt2", args.output / "raster.pt2"
-    torch.export.save(online_export, online_path)
-    torch.export.save(raster_export, raster_path)
+    _save_exported_program06(online_export, online_path)
+    _save_exported_program06(raster_export, raster_path)
     report = {
         "checkpoint": str(args.checkpoint), "adapter_checkpoint": str(args.adapter_checkpoint),
         "adapter_architecture": str(adapter_payload["adapter_architecture"]),
@@ -149,8 +221,8 @@ def main() -> None:
             report["litert"] = {"converted": False, "reason": "litert_torch_not_installed"}
         else:
             report["litert"] = {
-                "online": _convert_litert(online, online_inputs[0], args.output / "online.tflite"),
-                "raster": _convert_litert(raster, raster_inputs[0], args.output / "raster.tflite"),
+                "online": _convert_litert(online, online_inputs, args.output / "online.tflite"),
+                "raster": _convert_litert(raster, raster_inputs, args.output / "raster.tflite"),
             }
     report["torch_export_gate_passed"] = all(
         bool(report["torch_export"][name]["gate_passed"]) for name in ("online", "raster")