Add composite CPU latency and memory proxy

MANIFEST.json

@@ -1,11 +1,11 @@
 {
-    "schema":  "aiflow-hf-research-snapshot-v10",
-    "generated_at":  "2026-07-23T20:20:53.2135890Z",
+    "schema":  "aiflow-hf-research-snapshot-v11",
+    "generated_at":  "2026-07-23T20:26:01.9922255Z",
     "track":  "R_noncommercial_plus_rejected_P_proxy",
     "product_validation":  false,
     "public_release":  true,
     "contains_raw_dataset":  false,
-    "tests":  "285 passed",
+    "tests":  "287 passed",
     "retracted_paths":  [
                             "models/auxiliary/boundary_auxiliary_head.pt",
                             "models/auxiliary/seed17/boundary_joint_delta.pt",
@@ -30,19 +30,19 @@
                   },
                   {
                       "path":  "reports/RESEARCH_REPORT.md",
-                      "bytes":  34631,
-                      "sha256":  "46946707f85cd6860161e301c21970b885dfd742c122eb9d9f0f44c3173c208b"
-                  },
-                  {
-                      "path":  "scripts/evaluate_crohme_tray_joint_selector.py",
-                      "bytes":  11671,
-                      "sha256":  "58c175a95cdeb9fc764fb96594a76888122b8f118f2f26da1283e743b1a7cd04"
+                      "bytes":  35675,
+                      "sha256":  "e9064d37996ceb714d6ccea391b6cc00be05653a8b5190ee7fb3affda7f08d87"
                   },
                   {
                       "path":  "scripts/build_math_ink_06_litert_colab_bundle.py",
                       "bytes":  6862,
                       "sha256":  "f008635ab59344005f65f1394f56275e4c92610cd25781a2d95243d060010718"
                   },
+                  {
+                      "path":  "scripts/benchmark_math_ink_06_composite.py",
+                      "bytes":  7618,
+                      "sha256":  "6c5347fe68114116e0dfd29e55e3682f47b3ce42af1412009ab7883e2076887c"
+                  },
                   {
                       "path":  "scripts/audit_math_ink_06_local_baseline_overmerge.py",
                       "bytes":  11900,
@@ -88,6 +88,11 @@
                       "bytes":  33727,
                       "sha256":  "c3dfd328d9ff0a55eed2cb74657ad87e5e164968fddd7fae8c8fd92e137a4557"
                   },
+                  {
+                      "path":  "scripts/evaluate_crohme_tray_joint_selector.py",
+                      "bytes":  11671,
+                      "sha256":  "58c175a95cdeb9fc764fb96594a76888122b8f118f2f26da1283e743b1a7cd04"
+                  },
                   {
                       "path":  "src/math_ink_06.py",
                       "bytes":  58690,
@@ -215,13 +220,13 @@
                   },
                   {
                       "path":  "MODEL_INDEX.json",
-                      "bytes":  2473,
-                      "sha256":  "8926674c956664b98d2fc1519e75355ad13efe57228d122ce208defebdfb90f1"
+                      "bytes":  2766,
+                      "sha256":  "68ab1d983dc5da5cf547ba34e59ecae3023b5a66b73df7ee94be15445e428ef4"
                   },
                   {
                       "path":  "README.md",
-                      "bytes":  11437,
-                      "sha256":  "5a160ef5bb7753ad36fb39f692aecb3171c243743d8fdd4f64def16a3714af42"
+                      "bytes":  11920,
+                      "sha256":  "24c767d993cd750aa85a30708748386d7ab89c133f4dd756d3424f66c6958a14"
                   },
                   {
                       "path":  "assets/stroke_encoding.svg",
@@ -238,6 +243,16 @@
                       "bytes":  824191,
                       "sha256":  "18aff472c1bdb25a519b1553364229c9869aed1f36a9d663cf88a1bc808fab8f"
                   },
+                  {
+                      "path":  "models/seed17/behavior_role_head.pt",
+                      "bytes":  75794,
+                      "sha256":  "57ced82a2600bd342bba68ba0636ecc3fc98c3dc461e26a132b5264d0a3d95bd"
+                  },
+                  {
+                      "path":  "reports/behavior_role_seed47.json",
+                      "bytes":  30091,
+                      "sha256":  "3d4dcbe2f2b0afb483f992dacd62aab1089bdc41d55cc259bcb7fc21b5294db0"
+                  },
                   {
                       "path":  "reports/behavior_role_seed31.json",
                       "bytes":  45412,
@@ -249,34 +264,34 @@
                       "sha256":  "e2072fec1b1065777f0e9218c48d27f096b22f5545e7cc8d9962939282719561"
                   },
                   {
-                      "path":  "reports/behavior_role_3seed_summary.json",
-                      "bytes":  3567,
-                      "sha256":  "b16049218d2d006d952f59f55642fa9f3fbfbc3cc3a730fbd7b075f507c3b251"
-                  },
-                  {
-                      "path":  "reports/behavior_role_seed47.json",
-                      "bytes":  30091,
-                      "sha256":  "3d4dcbe2f2b0afb483f992dacd62aab1089bdc41d55cc259bcb7fc21b5294db0"
+                      "path":  "reports/boundary_behavior_guard_report.json",
+                      "bytes":  72053,
+                      "sha256":  "f55a230a9f0892b9676c71db46e9e527862b3d8da776dc43db767ce7a0f6a5c7"
                   },
                   {
                       "path":  "reports/local_baseline_guard_report.json",
                       "bytes":  24052,
                       "sha256":  "8f209b2890245f01f437c26e7556e19fee76c057d449c9382e2d64a70a3764d6"
                   },
+                  {
+                      "path":  "reports/composite_cpu_benchmark.json",
+                      "bytes":  2948,
+                      "sha256":  "1e9c75082cde0ac2c8ef303f2ff8870b3d2c772ad63027226e632b56f02ed9aa"
+                  },
                   {
                       "path":  "reports/component_competition_report.json",
                       "bytes":  22284,
                       "sha256":  "4d3db5f71e5008381dd072f2c10aafb22e08f96e23f2a32f82a3a46abd9c70b9"
                   },
                   {
-                      "path":  "reports/boundary_behavior_guard_report.json",
-                      "bytes":  72053,
-                      "sha256":  "f55a230a9f0892b9676c71db46e9e527862b3d8da776dc43db767ce7a0f6a5c7"
+                      "path":  "reports/behavior_role_3seed_summary.json",
+                      "bytes":  3567,
+                      "sha256":  "b16049218d2d006d952f59f55642fa9f3fbfbc3cc3a730fbd7b075f507c3b251"
                   },
                   {
-                      "path":  "models/seed47/online_adapter.pt",
-                      "bytes":  2876038,
-                      "sha256":  "46394eeebad3d27d3178f78e693e160700ba3f454d79044714b835bc51c07056"
+                      "path":  "models/seed31/behavior_role_head.pt",
+                      "bytes":  75794,
+                      "sha256":  "b7c22cf49320b90a6fd103d39cc3ff08ea7734ab83275aad66e2b529b749d9ae"
                   },
                   {
                       "path":  "models/seed31/base_378.pt",
@@ -289,14 +304,14 @@
                       "sha256":  "8ea224d7bc107a4c8b6cb91c9de6598000dae463d3e6d39480c081ea728cc87f"
                   },
                   {
-                      "path":  "models/seed17/behavior_role_head.pt",
-                      "bytes":  75794,
-                      "sha256":  "57ced82a2600bd342bba68ba0636ecc3fc98c3dc461e26a132b5264d0a3d95bd"
+                      "path":  "models/seed31/online_adapter.pt",
+                      "bytes":  2875654,
+                      "sha256":  "fedbbab64cea6755e76e0633367ee45fbcc1d5726bb9dfdb7c5fe4bbc968bdf2"
                   },
                   {
-                      "path":  "models/seed31/behavior_role_head.pt",
-                      "bytes":  75794,
-                      "sha256":  "b7c22cf49320b90a6fd103d39cc3ff08ea7734ab83275aad66e2b529b749d9ae"
+                      "path":  "models/seed47/online_adapter.pt",
+                      "bytes":  2876038,
+                      "sha256":  "46394eeebad3d27d3178f78e693e160700ba3f454d79044714b835bc51c07056"
                   },
                   {
                       "path":  "models/seed47/behavior_role_head.pt",
@@ -307,11 +322,6 @@
                       "path":  "models/seed47/base_378.pt",
                       "bytes":  8638290,
                       "sha256":  "4131b21f6b00471c1c5a82aee4a2da45e599e42dc742fc1dd0dd8afb970f8939"
-                  },
-                  {
-                      "path":  "models/seed31/online_adapter.pt",
-                      "bytes":  2875654,
-                      "sha256":  "fedbbab64cea6755e76e0633367ee45fbcc1d5726bb9dfdb7c5fe4bbc968bdf2"
                   }
               ]
 }

MODEL_INDEX.json

@@ -68,6 +68,14 @@
     "representative_samples": 76,
     "litert_flatbuffer": false
   },
+  "composite_cpu_proxy": {
+    "report": "reports/composite_cpu_benchmark.json",
+    "online_p95_ms_range": [7.77, 10.2],
+    "raster_p95_ms_range": [19.18, 23.57],
+    "model_state_bytes": 8947108,
+    "model_state_plus_inference_growth_bytes": 26301860,
+    "android_validation": false
+  },
   "release_state": {
     "track": "R_noncommercial_only",
     "product_validation": false,

README.md

@@ -207,6 +207,10 @@ Export 그래프는 이제 base-only가 아니라 online/raster modality adapter
 
 `.pt2`는 Android용 `.tflite`가 아니다. LiteRT Torch 0.9.1 변환과 Android runtime parity는 아직 완료되지 않았으므로 `litert_exported=false`를 유지한다.
 
+### CPU latency·memory proxy
+
+Windows PyTorch CPU의 실제 대표 입력 76개 측정에서 online p95는 7.77~10.20ms, raster p95는 19.18~23.57ms였다. Tensor state는 8.95MB, 모델 로드 후 inference RSS 증가분을 합친 구조 proxy는 26.30MB다. 전체 Python process RSS 231.5MB는 PyTorch runtime을 포함하므로 Android LiteRT memory 근거가 아니다.
+
 ## 출력 범위
 
 의도한 모바일 API:
@@ -259,6 +263,7 @@ 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)
+- Composite CPU benchmark: [`reports/composite_cpu_benchmark.json`](reports/composite_cpu_benchmark.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)

reports/RESEARCH_REPORT.md

@@ -413,6 +413,18 @@ Linux/Colab 변환을 위해 label이 없는 실제 HWRT-derived online/raster
 
 Python 공개 런타임의 `MathInk06Engine`도 base-only였으므로 `adapter_checkpoint`를 공식 생성자 인자로 추가했다. 엔진은 base→adapter `shared_state_dict`→dual modality adapter 순서로 구성하고 online은 `adapter.online`, raster virtual trajectory는 `adapter.raster`를 거친다. 실제 seed-17 representative 입력에서 엔진과 export wrapper의 online/raster 최대 logit 오차는 모두 0.0이었다. Adapter를 생략한 호출은 호환용 base-only 경로로만 남긴다.
 
+### Composite CPU latency·memory proxy
+
+정정된 Python runtime을 실제 representative 76개로 Windows CPU에서 측정했다. 각 thread 조건은 같은 입력과 output checksum을 사용했다.
+
+| Intra-op threads | Online p95 | Raster p95 |
+|---:|---:|---:|
+| 1 | 10.20ms | 20.65ms |
+| 2 | 8.51ms | 23.57ms |
+| 4 | 7.77ms | 19.18ms |
+
+세 조건 모두 online 50ms, raster 200ms software proxy를 통과했다. Model과 composite adapter의 중복 제거 tensor state는 8,947,108 bytes다. 모델 로드 후 관측한 inference RSS 증가분은 17,354,752 bytes이고 두 값을 합친 구조 proxy는 26,301,860 bytes로 100MiB 안이다. 반면 전체 Python process RSS는 최대 231,530,496 bytes였는데 이는 CPython·PyTorch runtime을 포함하므로 Android peak memory와 직접 비교하지 않는다. 실제 Android LiteRT 3-tier latency·memory·battery·delegate 검증은 계속 필요하다.
+
 ### 실제 P 연속식 데이터 계약
 
 실제 데이터가 도착하기 전에 `AIFlow P Formula v1` JSON Schema와 fail-closed preflight를 추가했다. 각 formula는 다음 근거를 모두 가져야 한다.
@@ -424,7 +436,7 @@ Python 공개 런타임의 `MathInk06Engine`도 base-only였으므로 `adapter_c
 - 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 285개가 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 287개가 통과했다.
 
 ## 산출물
 
@@ -450,6 +462,7 @@ Origin·writer·device·source가 둘 이상의 split에 나타나면 제품 평
 - `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/benchmark_math_ink_06_composite.py`
 - `scripts/preflight_math_ink_06_p_formula.py`
 - `scripts/analyze_crohme_lattice_failures.py`
 - `tests/test_behavior_context06.py`
@@ -458,6 +471,7 @@ Origin·writer·device·source가 둘 이상의 split에 나타나면 제품 평
 - `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/runs/math_ink_06_composite_cpu_benchmark_20260724/report.json`
 - `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`

reports/composite_cpu_benchmark.json

@@ -0,0 +1,92 @@
+{
+  "schema": "aiflow-math-ink-06-composite-cpu-benchmark-v1",
+  "generated_at": "2026-07-23T20:24:41.860993+00:00",
+  "torch_version": "2.5.1+cpu",
+  "platform": "win32",
+  "model_version": "aiflow-math-ink-0.6-federated-online1+aiflow-math-ink-0.6-skeleton-adapter1",
+  "model_state_bytes": 8947108,
+  "baseline_process_rss_bytes": 214175744,
+  "maximum_observed_process_rss_bytes": 231530496,
+  "inference_rss_growth_bytes": 17354752,
+  "model_state_plus_inference_growth_bytes": 26301860,
+  "memory_proxy_gate_le_100mib": true,
+  "rows": [
+    {
+      "threads": 1,
+      "online": {
+        "samples": 76,
+        "mean_ms": 6.648381579189414,
+        "p50_ms": 6.038199993781745,
+        "p95_ms": 10.202200006460771,
+        "maximum_ms": 10.744899991550483,
+        "output_checksum": 352945188,
+        "observed_process_rss_bytes": 219648000
+      },
+      "raster": {
+        "samples": 76,
+        "mean_ms": 19.39044473711922,
+        "p50_ms": 19.06099999905564,
+        "p95_ms": 20.654299994930625,
+        "maximum_ms": 28.47059999476187,
+        "output_checksum": 1642561801,
+        "observed_process_rss_bytes": 226373632
+      },
+      "proxy_gates": {
+        "online_p95_le_50ms": true,
+        "raster_p95_le_200ms": true
+      }
+    },
+    {
+      "threads": 2,
+      "online": {
+        "samples": 76,
+        "mean_ms": 6.068777630724454,
+        "p50_ms": 5.20410000171978,
+        "p95_ms": 8.508300001267344,
+        "maximum_ms": 8.986200002254918,
+        "output_checksum": 352945188,
+        "observed_process_rss_bytes": 226631680
+      },
+      "raster": {
+        "samples": 76,
+        "mean_ms": 16.508685525255522,
+        "p50_ms": 14.63520000106655,
+        "p95_ms": 23.574999999254942,
+        "maximum_ms": 24.597199997515418,
+        "output_checksum": 1642561801,
+        "observed_process_rss_bytes": 230453248
+      },
+      "proxy_gates": {
+        "online_p95_le_50ms": true,
+        "raster_p95_le_200ms": true
+      }
+    },
+    {
+      "threads": 4,
+      "online": {
+        "samples": 76,
+        "mean_ms": 6.707876315646756,
+        "p50_ms": 7.187800001702271,
+        "p95_ms": 7.77469998865854,
+        "maximum_ms": 8.105300003080629,
+        "output_checksum": 352945188,
+        "observed_process_rss_bytes": 230817792
+      },
+      "raster": {
+        "samples": 76,
+        "mean_ms": 17.300657895822567,
+        "p50_ms": 17.91409999714233,
+        "p95_ms": 19.176700006937608,
+        "maximum_ms": 20.39449999574572,
+        "output_checksum": 1642561801,
+        "observed_process_rss_bytes": 231530496
+      },
+      "proxy_gates": {
+        "online_p95_le_50ms": true,
+        "raster_p95_le_200ms": true
+      }
+    }
+  ],
+  "interpretation_limit": "Windows PyTorch CPU proxy이며 Android LiteRT·배터리·delegate 성능 판정이 아니다.",
+  "product_validation": false
+}

scripts/benchmark_math_ink_06_composite.py

@@ -0,0 +1,183 @@
+"""정정된 Math Ink 0.6 composite online/raster CPU 지연을 대표 입력으로 측정한다."""
+
+from __future__ import annotations
+
+import argparse
+from datetime import datetime, timezone
+import json
+import math
+from pathlib import Path
+import statistics
+import sys
+import time
+
+import torch
+
+PROJECT_ROOT = Path(__file__).parents[1]
+SOURCE_ROOT = PROJECT_ROOT / "src"
+for path in (PROJECT_ROOT, SOURCE_ROOT):
+    if str(path) not in sys.path:
+        sys.path.insert(0, str(path))
+
+from math_grid_drawer.research.math_ink_06 import MathInk06Engine
+from scripts.export_math_ink_06_litert import _load_representative_inputs06
+
+
+def percentile_nearest_rank06(values: list[float], percentile: float) -> float:
+    """필요 변수: 측정값·0~1 percentile. 작동 원리: 모바일 p95와 동일한 nearest-rank 값을 반환한다."""
+
+    if not values or not 0.0 <= percentile <= 1.0:
+        raise ValueError("percentile 입력이 유효하지 않습니다.")
+    ordered = sorted(values)
+    index = max(0, min(len(ordered) - 1, math.ceil(percentile * len(ordered)) - 1))
+    return float(ordered[index])
+
+
+def module_state_bytes06(*modules: torch.nn.Module) -> int:
+    """필요 변수: model·adapter module. 작동 원리: 중복 storage를 한 번만 세어 실제 tensor state bytes를 계산한다."""
+
+    seen: set[tuple[int, int]] = set()
+    total = 0
+    for module in modules:
+        for tensor in [*module.parameters(), *module.buffers()]:
+            storage = tensor.untyped_storage()
+            key = (storage.data_ptr(), storage.nbytes())
+            if key in seen:
+                continue
+            seen.add(key)
+            total += storage.nbytes()
+    return total
+
+
+def _rss_bytes06() -> int | None:
+    """필요 변수: 없음. 작동 원리: psutil이 있으면 현재 process RSS를 반환하고 없으면 명시적으로 결측 처리한다."""
+
+    try:
+        import psutil
+    except ImportError:
+        return None
+    return int(psutil.Process().memory_info().rss)
+
+
+def _measure06(callable_, inputs: list[tuple[torch.Tensor, ...]], warmup: int) -> dict:
+    """필요 변수: 고정 inference callable·대표 입력·warmup. 작동 원리: 표본별 wall latency와 output checksum을 측정한다."""
+
+    with torch.inference_mode():
+        for arguments in inputs[:max(1, min(warmup, len(inputs)))]:
+            callable_(*arguments)
+        latencies, checksum = [], 0
+        observed_rss = _rss_bytes06()
+        for arguments in inputs:
+            started = time.perf_counter()
+            output = callable_(*arguments)
+            latencies.append((time.perf_counter() - started) * 1000.0)
+            primary = output[0] if isinstance(output, tuple) else output
+            checksum = (checksum * 131 + int(primary.argmax(dim=-1)[0])) % 2_147_483_647
+            current_rss = _rss_bytes06()
+            if current_rss is not None:
+                observed_rss = max(observed_rss or 0, current_rss)
+    return {
+        "samples": len(latencies),
+        "mean_ms": statistics.fmean(latencies),
+        "p50_ms": percentile_nearest_rank06(latencies, 0.50),
+        "p95_ms": percentile_nearest_rank06(latencies, 0.95),
+        "maximum_ms": max(latencies),
+        "output_checksum": checksum,
+        "observed_process_rss_bytes": observed_rss,
+    }
+
+
+def main() -> None:
+    """필요 변수: composite artifact·대표 cache. 작동 원리: thread별 두 inference 경로를 독립 측정해 JSON으로 남긴다."""
+
+    parser = argparse.ArgumentParser(description="Benchmark Math Ink 0.6 composite CPU")
+    parser.add_argument("--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("--threads", type=int, action="append", default=None)
+    parser.add_argument("--samples", type=int, default=76)
+    parser.add_argument("--warmup", type=int, default=5)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    requested_threads = args.threads or [1, 2, 4]
+    if any(value <= 0 for value in requested_threads):
+        raise ValueError("CPU thread는 양수여야 합니다.")
+    torch.set_num_interop_threads(1)
+    engine = MathInk06Engine(
+        args.checkpoint, adapter_checkpoint=args.adapter_checkpoint, device="cpu",
+    )
+    online_inputs, raster_inputs = _load_representative_inputs06(args.representative_inputs)
+    online_inputs = online_inputs[:args.samples]
+    raster_inputs = raster_inputs[:args.samples]
+
+    def online_forward(sequence: torch.Tensor):
+        """필요 변수: canonical sequence. 작동 원리: 실제 runtime online composite branch를 호출한다."""
+
+        return engine.model.forward_online(engine.online_adapter(sequence))
+
+    def raster_forward(raster: torch.Tensor):
+        """필요 변수: raster. 작동 원리: 실제 runtime virtual stroke·raster adapter·fusion을 호출한다."""
+
+        output = engine._forward_raster_composite06(raster)
+        return engine.fuse_raster_output(output)[0]
+
+    baseline_rss = _rss_bytes06()
+    rows = []
+    for thread_count in requested_threads:
+        torch.set_num_threads(thread_count)
+        online = _measure06(online_forward, online_inputs, args.warmup)
+        raster = _measure06(raster_forward, raster_inputs, args.warmup)
+        rows.append({
+            "threads": thread_count,
+            "online": online,
+            "raster": raster,
+            "proxy_gates": {
+                "online_p95_le_50ms": online["p95_ms"] <= 50.0,
+                "raster_p95_le_200ms": raster["p95_ms"] <= 200.0,
+            },
+        })
+    observed_rss_values = [
+        int(metrics["observed_process_rss_bytes"])
+        for row in rows for metrics in (row["online"], row["raster"])
+        if metrics["observed_process_rss_bytes"] is not None
+    ]
+    maximum_observed_rss = max(observed_rss_values) if observed_rss_values else None
+    inference_rss_growth = (
+        max(0, maximum_observed_rss - baseline_rss)
+        if maximum_observed_rss is not None and baseline_rss is not None else None
+    )
+    state_bytes = module_state_bytes06(engine.model, engine.composite_adapter)
+    model_plus_inference = (
+        state_bytes + inference_rss_growth
+        if inference_rss_growth is not None else None
+    )
+    report = {
+        "schema": "aiflow-math-ink-06-composite-cpu-benchmark-v1",
+        "generated_at": datetime.now(timezone.utc).isoformat(),
+        "torch_version": torch.__version__,
+        "platform": sys.platform,
+        "model_version": engine.model_version,
+        "model_state_bytes": state_bytes,
+        "baseline_process_rss_bytes": baseline_rss,
+        "maximum_observed_process_rss_bytes": maximum_observed_rss,
+        "inference_rss_growth_bytes": inference_rss_growth,
+        "model_state_plus_inference_growth_bytes": model_plus_inference,
+        "memory_proxy_gate_le_100mib": (
+            model_plus_inference <= 100 * 1024 * 1024
+            if model_plus_inference is not None else None
+        ),
+        "rows": rows,
+        "interpretation_limit": (
+            "Windows PyTorch CPU proxy이며 Android LiteRT·배터리·delegate 성능 판정이 아니다."
+        ),
+        "product_validation": False,
+    }
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.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()