{
 "meta": {
  "date_utc": "2026-10-12T04:46:06Z",
  "release": {
   "tag": "v0.2.4",
   "url": "https://github.com/gist-rs/reflex/releases/tag/v0.2.4",
   "archives": {
    "reflex-v0.2.4-aarch64-apple-darwin.tar.gz": 1880244,
    "reflex-v0.2.4-aarch64-unknown-linux-musl.tar.gz": 1925853,
    "reflex-v0.2.4-x86_64-apple-darwin.tar.gz": 2051317,
    "reflex-v0.2.4-x86_64-pc-windows-gnu.zip": 2048162,
    "reflex-v0.2.4-x86_64-unknown-linux-musl.tar.gz": 2117218
   }
  },
  "law": "every byte count is measured (release archive unpack, HF tree API, local stat) or a recorded on-host measurement with provenance \u2014 never hand-typed",
  "order": "ascending total (engine + model)"
 },
 "candidates": [
  {
   "key": "wasm_heads",
   "name": "Reflex \u00b7 in-browser (wasm heads)",
   "framework": "Rust \u2192 wasm32 (wasm-opt -Oz), zero deps",
   "engine_kind": "rust",
   "engine_bytes": 92057,
   "engine_what": "arena_head.wasm \u2014 the engine's fitted game heads as one module",
   "engine_provenance": {
    "source": "local file",
    "detail": "assets/arena_head.wasm",
    "parts": [
     {
      "source": "local file",
      "what": "assets/arena_head.wasm"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "arena_head.wasm",
     "kind": "module",
     "bytes": 92057,
     "src": 0
    }
   ],
   "model_bytes": 0,
   "model_what": "none \u2014 the fitted heads are compiled into the module",
   "model_provenance": null,
   "targets": [
    "browser tab",
    "edge worker (1 MB class)"
   ],
   "note": "plays Tetris / Flappy / three-lanes in-tab with no engine install \u2014 parity-proven against the recorded engine play before it moves a piece"
  },
  {
   "key": "reflex_native",
   "name": "Reflex \u00b7 native binary (modelless)",
   "framework": "one static Rust binary (dist profile, stripped, fat LTO)",
   "engine_kind": "rust",
   "engine_bytes": 4381977,
   "engine_what": "installed binary + THIRD_PARTY_LICENSES.md (unpacked from the aarch64-apple-darwin archive)",
   "engine_provenance": {
    "source": "measured: unpacked release archive",
    "detail": "v0.2.4 aarch64-apple-darwin, downloaded + stat'd at publish time",
    "parts": [
     {
      "source": "measured: unpacked release archive",
      "what": "v0.2.4 aarch64-apple-darwin, downloaded + stat'd at publish time"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "reflex binary",
     "kind": "binary",
     "bytes": 4329680,
     "src": 0
    },
    {
     "label": "licenses + archive metadata (5 files)",
     "kind": "license",
     "bytes": 52297,
     "src": 0
    }
   ],
   "model_bytes": 0,
   "model_what": "none \u2014 the corpus ships inside the binary; the game heads boot-fit from embedded fixtures",
   "model_provenance": null,
   "targets": [
    "macOS",
    "Linux (musl, static)",
    "Windows",
    "container"
   ],
   "note": "archive download 1.88\u20132.12 MB across all 5 release targets \u2014 the same binary serves every lane"
  },
  {
   "key": "instinct_hybrid",
   "name": "Instinct \u00b7 trained specialists",
   "framework": "one serve binary + the Reflex half's dataset seats + BLAKE3-pinned specialist files (the hosted serving posture)",
   "engine_kind": "rust",
   "engine_bytes": 22596268,
   "engine_what": "the serve binary + the six t20k dataset suites (the Reflex half's corpora and question seats)",
   "engine_provenance": {
    "source": "recorded measurement (sum)",
    "detail": "1,787,744 B (the riir-instinct `serve` release binary (aarch64-apple-darwin) \u2014 the hosted hybrid lane's serving edge, 2026-09-27 on m3-max-metal) + 20,808,524 B (the six t20k dataset suites the serving posture stages (ag_news 6,324,897 \u00b7 emotion 2,719,177 \u00b7 sst5 1,799,678 \u00b7 massive_intent_en 2,516,925 \u00b7 banking77 2,157,368 \u00b7 xnli_en 5,290,479) \u2014 the Reflex half's corpora and question seats, 2026-09-27 on m3-max-metal)",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "serve binary",
      "what": "the riir-instinct `serve` release binary (aarch64-apple-darwin) \u2014 the hosted hybrid lane's serving edge",
      "date_utc": "2026-09-27",
      "host": "m3-max-metal",
      "how": "lstat over ../riir-instinct/target/release/serve (cargo build --release --bin serve)"
     },
     {
      "source": "recorded measurement",
      "label": "six t20k dataset suites",
      "what": "the six t20k dataset suites the serving posture stages (ag_news 6,324,897 \u00b7 emotion 2,719,177 \u00b7 sst5 1,799,678 \u00b7 massive_intent_en 2,516,925 \u00b7 banking77 2,157,368 \u00b7 xnli_en 5,290,479) \u2014 the Reflex half's corpora and question seats",
      "date_utc": "2026-09-27",
      "host": "m3-max-metal",
      "how": "sum of lstat sizes over ../riir-reflex/.raw/datasets_t20k/{ag_news,emotion,sst5,massive_intent_en,banking77,xnli_en} (the deploy.yaml files: rows)"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "serve binary",
     "kind": "binary",
     "bytes": 1787744,
     "src": 0
    },
    {
     "label": "ag_news dataset",
     "kind": "dataset",
     "bytes": 6324897,
     "src": 1
    },
    {
     "label": "emotion dataset",
     "kind": "dataset",
     "bytes": 2719177,
     "src": 1
    },
    {
     "label": "sst5 dataset",
     "kind": "dataset",
     "bytes": 1799678,
     "src": 1
    },
    {
     "label": "massive_intent_en dataset",
     "kind": "dataset",
     "bytes": 2516925,
     "src": 1
    },
    {
     "label": "banking77 dataset",
     "kind": "dataset",
     "bytes": 2157368,
     "src": 1
    },
    {
     "label": "xnli_en dataset",
     "kind": "dataset",
     "bytes": 5290479,
     "src": 1
    }
   ],
   "model_bytes": 20320298,
   "model_what": "the six winner files (each pinned by a BLAKE3 digest in the manifest, checked at load \u2014 the signed HOSTED-ONLY vessel is not yet the serving path)",
   "model_provenance": {
    "source": "recorded measurement",
    "detail": "the six winner files (BLAKE3-pinned in the manifest): ag_news 524,381 \u00b7 emotion 786,545 \u00b7 sst5 655,460 \u00b7 massive_intent_en 7,865,749 \u00b7 banking77 10,094,864 \u00b7 xnli_en 393,299 \u2014 measured 2026-09-27 on m3-max-metal: sum of lstat sizes over ../riir-train/data/instinct_specialists/{ag_news,emotion,sst5,massive_intent_en,banking77,xnli_en}_winner_v1.bin (the deploy.yaml files: rows)",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "six specialist winner files",
      "what": "the six winner files (BLAKE3-pinned in the manifest): ag_news 524,381 \u00b7 emotion 786,545 \u00b7 sst5 655,460 \u00b7 massive_intent_en 7,865,749 \u00b7 banking77 10,094,864 \u00b7 xnli_en 393,299",
      "date_utc": "2026-09-27",
      "host": "m3-max-metal",
      "how": "sum of lstat sizes over ../riir-train/data/instinct_specialists/{ag_news,emotion,sst5,massive_intent_en,banking77,xnli_en}_winner_v1.bin (the deploy.yaml files: rows)"
     }
    ]
   },
   "model_stack": [
    {
     "label": "ag_news specialist",
     "kind": "specialist",
     "bytes": 524381,
     "src": 0
    },
    {
     "label": "emotion specialist",
     "kind": "specialist",
     "bytes": 786545,
     "src": 0
    },
    {
     "label": "sst5 specialist",
     "kind": "specialist",
     "bytes": 655460,
     "src": 0
    },
    {
     "label": "massive_intent_en specialist",
     "kind": "specialist",
     "bytes": 7865749,
     "src": 0
    },
    {
     "label": "banking77 specialist",
     "kind": "specialist",
     "bytes": 10094864,
     "src": 0
    },
    {
     "label": "xnli_en specialist",
     "kind": "specialist",
     "bytes": 393299,
     "src": 0
    }
   ],
   "targets": [
    "container (cf-container)",
    "hosted serving"
   ],
   "note": "the trained sibling lane: the specialists serve six text suites; game spots answer through its Reflex half. The three-product naming is Reflex / Instinct / Rethink \u2014 Instinct is the open teaching lane (052); Rethink is the private moat"
  },
  {
   "key": "rethink_encoder",
   "name": "Rethink \u00b7 encoder arm",
   "framework": "one serve binary (GPU host) + the laya-riir encoder resident from boot + the Reflex half's dataset seats",
   "engine_kind": "rust",
   "engine_bytes": 28748684,
   "engine_what": "the serve binary (the encoder lane's GPU-host build) + the six t20k dataset suites (the seats the ENC lanes boot from)",
   "engine_provenance": {
    "source": "recorded measurement (sum)",
    "detail": "7,940,160 B (the riir-instinct `serve` release binary built with --features serve-encoder-metal (aarch64-apple-darwin) \u2014 the Rethink encoder lane's GPU-host serving edge (the encoder lane + the laya-riir Metal backend compiled in; the CUDA twin is the same size class), 2026-10-01 on m3-max-metal) + 20,808,524 B (the six t20k dataset suites the serving posture stages (ag_news 6,324,897 \u00b7 emotion 2,719,177 \u00b7 sst5 1,799,678 \u00b7 massive_intent_en 2,516,925 \u00b7 banking77 2,157,368 \u00b7 xnli_en 5,290,479) \u2014 the Reflex half's corpora and question seats, 2026-09-27 on m3-max-metal)",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "serve binary \u00b7 GPU-host build",
      "what": "the riir-instinct `serve` release binary built with --features serve-encoder-metal (aarch64-apple-darwin) \u2014 the Rethink encoder lane's GPU-host serving edge (the encoder lane + the laya-riir Metal backend compiled in; the CUDA twin is the same size class)",
      "date_utc": "2026-10-01",
      "host": "m3-max-metal",
      "how": "lstat over ../riir-instinct/target/release/serve (cargo build --release --bin serve --features serve-encoder-metal; the binary's build stamp carries arena-laya, serve-encoder, serve-encoder-metal)"
     },
     {
      "source": "recorded measurement",
      "label": "six t20k dataset suites",
      "what": "the six t20k dataset suites the serving posture stages (ag_news 6,324,897 \u00b7 emotion 2,719,177 \u00b7 sst5 1,799,678 \u00b7 massive_intent_en 2,516,925 \u00b7 banking77 2,157,368 \u00b7 xnli_en 5,290,479) \u2014 the Reflex half's corpora and question seats",
      "date_utc": "2026-09-27",
      "host": "m3-max-metal",
      "how": "sum of lstat sizes over ../riir-reflex/.raw/datasets_t20k/{ag_news,emotion,sst5,massive_intent_en,banking77,xnli_en} (the deploy.yaml files: rows)"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "serve binary \u00b7 GPU-host build",
     "kind": "binary",
     "bytes": 7940160,
     "src": 0
    },
    {
     "label": "ag_news dataset",
     "kind": "dataset",
     "bytes": 6324897,
     "src": 1
    },
    {
     "label": "emotion dataset",
     "kind": "dataset",
     "bytes": 2719177,
     "src": 1
    },
    {
     "label": "sst5 dataset",
     "kind": "dataset",
     "bytes": 1799678,
     "src": 1
    },
    {
     "label": "massive_intent_en dataset",
     "kind": "dataset",
     "bytes": 2516925,
     "src": 1
    },
    {
     "label": "banking77 dataset",
     "kind": "dataset",
     "bytes": 2157368,
     "src": 1
    },
    {
     "label": "xnli_en dataset",
     "kind": "dataset",
     "bytes": 5290479,
     "src": 1
    }
   ],
   "model_bytes": 455724111,
   "model_what": "the laya-english checkpoint as the derived Q8_0 artifact (the adopted encoding \u2014 LAYA_WEIGHTS_VARIANT=q8; 53.1% of the F16 file) + the three locked NLEH heads (sst5 \u00b7 xnli_en \u00b7 ag_news v1)",
   "model_provenance": {
    "source": "sum of measured sources",
    "detail": "447,729,135 B (recorded measurement: the laya-english checkpoint as the derived Q8_0 artifact (the ADOPTED serving encoding \u2014 LAYA_WEIGHTS_VARIANT=q8, riir-infer plan 616 Phases 1\u20132; 53.1% of the 842,609,210-byte F16 file) \u2014 the weights the Rethink encoder serving lane loads resident from boot \u2014 measured 2026-10-02 on m3-max-metal: lstat over ~/.cache/riir-reflex/laya/english/derived/model.q8.safetensors (stat -f %z \u2192 447729135); the model.q8.safetensors.blake3 sidecar sits beside it \u2014 digest NOT re-verified at record time (no blake3 CLI on the box)) + 7,994,976 B (recorded measurement: the three locked NLEH encoder heads the serving lane serves (english-checkpoint paired): sst5 t6_s0.bin 3,197,988 \u00b7 xnli_en_encoder_v1.bin 2,131,996 \u00b7 ag_news_encoder_v1.bin 2,664,992. The typed v2 head (typed_encoder_v2.bin, 1,066,004 \u2014 same dir) is NOT in this posture: it is trained over the TYPED checkpoint (encoder_serve hardcodes Checkpoint::English; 'a head trained on the typed cache is meaningless over the english encoder') and rides the arena's record-only measurement lane \u2014 measured 2026-10-01 on m3-max-metal: sum of lstat sizes over ../riir-train/.raw/t599/{t6_s0.bin,xnli_en_encoder_v1.bin,ag_news_encoder_v1.bin} (the .blake3 sidecars excluded))",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "laya-english encoder \u00b7 Q8_0",
      "what": "the laya-english checkpoint as the derived Q8_0 artifact (the ADOPTED serving encoding \u2014 LAYA_WEIGHTS_VARIANT=q8, riir-infer plan 616 Phases 1\u20132; 53.1% of the 842,609,210-byte F16 file) \u2014 the weights the Rethink encoder serving lane loads resident from boot",
      "date_utc": "2026-10-02",
      "host": "m3-max-metal",
      "how": "lstat over ~/.cache/riir-reflex/laya/english/derived/model.q8.safetensors (stat -f %z \u2192 447729135); the model.q8.safetensors.blake3 sidecar sits beside it \u2014 digest NOT re-verified at record time (no blake3 CLI on the box)"
     },
     {
      "source": "recorded measurement",
      "label": "three locked encoder heads",
      "what": "the three locked NLEH encoder heads the serving lane serves (english-checkpoint paired): sst5 t6_s0.bin 3,197,988 \u00b7 xnli_en_encoder_v1.bin 2,131,996 \u00b7 ag_news_encoder_v1.bin 2,664,992. The typed v2 head (typed_encoder_v2.bin, 1,066,004 \u2014 same dir) is NOT in this posture: it is trained over the TYPED checkpoint (encoder_serve hardcodes Checkpoint::English; 'a head trained on the typed cache is meaningless over the english encoder') and rides the arena's record-only measurement lane",
      "date_utc": "2026-10-01",
      "host": "m3-max-metal",
      "how": "sum of lstat sizes over ../riir-train/.raw/t599/{t6_s0.bin,xnli_en_encoder_v1.bin,ag_news_encoder_v1.bin} (the .blake3 sidecars excluded)"
     }
    ]
   },
   "model_stack": [
    {
     "label": "laya-english encoder \u00b7 Q8_0",
     "kind": "encoder",
     "bytes": 447729135,
     "src": 0
    },
    {
     "label": "sst5 head (t6_s0)",
     "kind": "head",
     "bytes": 3197988,
     "src": 1
    },
    {
     "label": "xnli_en head",
     "kind": "head",
     "bytes": 2131996,
     "src": 1
    },
    {
     "label": "ag_news head",
     "kind": "head",
     "bytes": 2664992,
     "src": 1
    }
   ],
   "targets": [
    "GPU host (Metal/CUDA)",
    "record-only today"
   ],
   "note": "the adopted q8 serving posture (riir-infer plan 616 Phases 1\u20132): 348.8 MiB device-resident q8 weights vs 1,654.9 MiB at the widened-F32 posture (4.74\u00d7) and 1,492 MiB whole-process RSS vs 4,825 MiB at the F16 host-widen posture \u2014 measured 2026-10-02 on m3-max-metal (AC). the lane is record-only today (GPU-host targets: the encoder class is refused at the CPU-only deploy shape, so nothing ships until a GPU serving deploy exists) \u2014 the serving lane loads the english checkpoint only; the typed cell is record-only (its v2 head is paired to the typed checkpoint through the arena's measurement lane, so that checkpoint + the typed full-pool corpus are recorded separately, not part of this posture) and serves only after a head retrained over english encodes or a typed serve lane; pre-adoption reference: the F16 english checkpoint tree was 848,195,504 B on HF \u2014 the row's model bytes before the q8 adoption"
  },
  {
   "key": "reflex_laya_typed",
   "name": "Reflex + laya \u00b7 typed",
   "framework": "the same reflex binary + the laya-riir lane (runtime download)",
   "engine_kind": "rust",
   "engine_bytes": 4381977,
   "engine_what": "the same installed binary",
   "engine_provenance": {
    "source": "measured: unpacked release archive",
    "detail": "v0.2.4 aarch64-apple-darwin, downloaded + stat'd at publish time",
    "parts": [
     {
      "source": "measured: unpacked release archive",
      "what": "v0.2.4 aarch64-apple-darwin, downloaded + stat'd at publish time"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "reflex binary",
     "kind": "binary",
     "bytes": 4329680,
     "src": 0
    },
    {
     "label": "licenses + archive metadata (5 files)",
     "kind": "license",
     "bytes": 52297,
     "src": 0
    }
   ],
   "model_bytes": 846195716,
   "model_what": "the typed-decisions specialist checkpoint (SHA-256-pinned, downloaded at first boot)",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "convaiinnovations/laya \u00b7 typed-decisions/ subtree sum",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "convaiinnovations/laya \u00b7 typed-decisions/ subtree sum"
     }
    ]
   },
   "model_stack": [
    {
     "label": "typed-decisions checkpoint",
     "kind": "weights",
     "bytes": 846195716,
     "src": 0
    }
   ],
   "targets": [
    "macOS (Metal)",
    "Linux",
    "Windows (CUDA)",
    "container"
   ],
   "note": "the english base is the same size class (848 MB); multilingual 678 MB; all three checkpoints \u2248 2.37 GB"
  },
  {
   "key": "bekko68m",
   "name": "Bekko-SystemOne-v0 (68M)",
   "framework": "their BekkoSentenceTransformer runtime in a torch venv (same venv as the 400M row \u2014 the two sizes share one engine)",
   "engine_kind": "python",
   "engine_bytes": 661500299,
   "engine_what": "the bekko lane venv (python 3.12: torch + transformers + sentence-transformers \u2014 the card's runtime pins)",
   "engine_provenance": {
    "source": "recorded measurement",
    "detail": "the bekko lane venv (python 3.12: torch 2.10.0 + transformers 5.17.0 + sentence-transformers 6.1.0 + safetensors + tqdm + huggingface-hub \u2014 the Bench-103 repro pins) \u2014 measured 2026-10-02 on m3-max-metal: uv venv --python 3.12 .raw/bekko-env + `uv pip install torch==2.10.0 transformers==5.17.0 sentence-transformers==6.1.0 safetensors tqdm 'huggingface-hub>=1,<2'` (the Bench-103 repro block); sum of file sizes (660,436,171 over 25,866 files) + dir entries (1,064,128), the lstat convention of the sibling venv rows.",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "bekko venv",
      "what": "the bekko lane venv (python 3.12: torch 2.10.0 + transformers 5.17.0 + sentence-transformers 6.1.0 + safetensors + tqdm + huggingface-hub \u2014 the Bench-103 repro pins)",
      "date_utc": "2026-10-02",
      "host": "m3-max-metal",
      "how": "uv venv --python 3.12 .raw/bekko-env + `uv pip install torch==2.10.0 transformers==5.17.0 sentence-transformers==6.1.0 safetensors tqdm 'huggingface-hub>=1,<2'` (the Bench-103 repro block); sum of file sizes (660,436,171 over 25,866 files) + dir entries (1,064,128), the lstat convention of the sibling venv rows."
     }
    ]
   },
   "engine_stack": [
    {
     "label": "bekko venv",
     "kind": "runtime",
     "bytes": 661500299,
     "src": 0
    }
   ],
   "model_bytes": 476225680,
   "model_what": "their 68M model tree (fp32 safetensors + tokenizer + the browser ONNX export, the card's pinned release revision 6eb1bae2) \u2014 the tree a consumer downloads incl. onnx_browser/",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "hotchpotch/bekko-system-one-v0-68m",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "hotchpotch/bekko-system-one-v0-68m"
     }
    ]
   },
   "model_stack": [
    {
     "label": "bekko-68m model tree",
     "kind": "weights",
     "bytes": 476225680,
     "src": 0
    }
   ],
   "targets": [
    "python env",
    "CPU (FP32 reference posture)"
   ],
   "note": "the lane's PRIOR seat (reflex Bench 103/104 era), superseded by the 400M on the board (Bench 107) \u2014 kept because it stays reachable: reflex pins resolve BEKKO_MODEL=bekko-system-one-v0-68m to it by name; the 68M-teacher distill record (riir-train Issue 608) reproduces against it; MIT (verified 2026-10-02)"
  },
  {
   "key": "laya_python",
   "name": "laya \u00b7 python reference",
   "framework": "CPython + torch (MPS build) + transformers + the pinned reference checkout",
   "engine_kind": "python",
   "engine_bytes": 643625020,
   "engine_what": "the oracle's python import closure + the pinned .raw/laya checkout",
   "engine_provenance": {
    "source": "recorded measurement",
    "detail": "python runtime for the ORIGINAL torch reference oracle: the import closure of scripts/laya_python_lane.py (torch 392.4 MB MPS build, transformers 95.4 MB, sympy, numpy, tokenizers, ... = 637497404 B) + the pinned reference checkout .raw/laya (6127616 B) \u2014 measured 2026-09-25 on m3-max-metal: sum of lstat sizes over the oracle's imported site-packages dirs (homebrew python3.14) + `du` over .raw/laya",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "python runtime + reference checkout",
      "what": "python runtime for the ORIGINAL torch reference oracle: the import closure of scripts/laya_python_lane.py (torch 392.4 MB MPS build, transformers 95.4 MB, sympy, numpy, tokenizers, ... = 637497404 B) + the pinned reference checkout .raw/laya (6127616 B)",
      "date_utc": "2026-09-25",
      "host": "m3-max-metal",
      "how": "sum of lstat sizes over the oracle's imported site-packages dirs (homebrew python3.14) + `du` over .raw/laya"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "python runtime + reference checkout",
     "kind": "runtime",
     "bytes": 643625020,
     "src": 0
    }
   ],
   "model_bytes": 846195716,
   "model_what": "the same typed checkpoint, loaded by the original reference",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "convaiinnovations/laya \u00b7 typed-decisions/ subtree sum",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "convaiinnovations/laya \u00b7 typed-decisions/ subtree sum"
     }
    ]
   },
   "model_stack": [
    {
     "label": "typed-decisions checkpoint",
     "kind": "weights",
     "bytes": 846195716,
     "src": 0
    }
   ],
   "targets": [
    "python env",
    "GPU (MPS/CUDA)"
   ],
   "note": "measurement-only lane \u2014 the reference is never shipped; footprint measured on the bench host that runs it"
  },
  {
   "key": "openthai",
   "name": "OpenThai-SystemOne",
   "framework": "their FastAPI service in a torch venv (the Thai/English decision comparison lane)",
   "engine_kind": "python",
   "engine_bytes": 677613688,
   "engine_what": "their service venv (torch + transformers + fastapi/uvicorn + their openthai_systemone package)",
   "engine_provenance": {
    "source": "recorded measurement",
    "detail": "the openthai-systemone service venv (python 3.12: torch + transformers + numpy + pydantic + fastapi/uvicorn + their openthai_systemone package installed from source @ 5d04bcca) \u2014 measured 2026-09-28 on m3-max-metal: uv venv --python 3.12 + `uv pip install torch transformers numpy pydantic fastapi uvicorn` + `uv pip install -e` their openthai-systemone @ iapp-technology/openthai-systemone 5d04bcca (the Bench-074 pin); sum of file sizes (676,763,544) + dir entries (850,144) over the venv, matching the lstat convention of the sibling venv rows. The Bench-074 venv itself was removed with .raw after the record \u2014 re-created once to re-measure the same dep set.",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "openthai service venv",
      "what": "the openthai-systemone service venv (python 3.12: torch + transformers + numpy + pydantic + fastapi/uvicorn + their openthai_systemone package installed from source @ 5d04bcca)",
      "date_utc": "2026-09-28",
      "host": "m3-max-metal",
      "how": "uv venv --python 3.12 + `uv pip install torch transformers numpy pydantic fastapi uvicorn` + `uv pip install -e` their openthai-systemone @ iapp-technology/openthai-systemone 5d04bcca (the Bench-074 pin); sum of file sizes (676,763,544) + dir entries (850,144) over the venv, matching the lstat convention of the sibling venv rows. The Bench-074 venv itself was removed with .raw after the record \u2014 re-created once to re-measure the same dep set."
     }
    ]
   },
   "engine_stack": [
    {
     "label": "openthai service venv",
     "kind": "runtime",
     "bytes": 677613688,
     "src": 0
    }
   ],
   "model_bytes": 1540237483,
   "model_what": "their model tree (Qwen3.5-0.8B text tower + the 256-slot decision head, safetensors + tokenizer)",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "iapp/OpenThai-SystemOne",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "iapp/OpenThai-SystemOne"
     }
    ]
   },
   "model_stack": [
    {
     "label": "Qwen3.5-0.8B + decision head",
     "kind": "weights",
     "bytes": 1540237483,
     "src": 0
    }
   ],
   "targets": [
    "python env",
    "GPU (MPS/CUDA)"
   ],
   "note": "the bench board's Thai-capability lane (Apache-2.0) \u2014 served on loopback, measured by our harness"
  },
  {
   "key": "bekko",
   "name": "Bekko-SystemOne-v0 (400M)",
   "framework": "their BekkoSentenceTransformer runtime in a torch venv",
   "engine_kind": "python",
   "engine_bytes": 661500299,
   "engine_what": "the bekko lane venv (python 3.12: torch + transformers + sentence-transformers \u2014 the card's runtime pins)",
   "engine_provenance": {
    "source": "recorded measurement",
    "detail": "the bekko lane venv (python 3.12: torch 2.10.0 + transformers 5.17.0 + sentence-transformers 6.1.0 + safetensors + tqdm + huggingface-hub \u2014 the Bench-103 repro pins) \u2014 measured 2026-10-02 on m3-max-metal: uv venv --python 3.12 .raw/bekko-env + `uv pip install torch==2.10.0 transformers==5.17.0 sentence-transformers==6.1.0 safetensors tqdm 'huggingface-hub>=1,<2'` (the Bench-103 repro block); sum of file sizes (660,436,171 over 25,866 files) + dir entries (1,064,128), the lstat convention of the sibling venv rows.",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "bekko venv",
      "what": "the bekko lane venv (python 3.12: torch 2.10.0 + transformers 5.17.0 + sentence-transformers 6.1.0 + safetensors + tqdm + huggingface-hub \u2014 the Bench-103 repro pins)",
      "date_utc": "2026-10-02",
      "host": "m3-max-metal",
      "how": "uv venv --python 3.12 .raw/bekko-env + `uv pip install torch==2.10.0 transformers==5.17.0 sentence-transformers==6.1.0 safetensors tqdm 'huggingface-hub>=1,<2'` (the Bench-103 repro block); sum of file sizes (660,436,171 over 25,866 files) + dir entries (1,064,128), the lstat convention of the sibling venv rows."
     }
    ]
   },
   "engine_stack": [
    {
     "label": "bekko venv",
     "kind": "runtime",
     "bytes": 661500299,
     "src": 0
    }
   ],
   "model_bytes": 3012643664,
   "model_what": "their 400M model tree (fp32 safetensors + tokenizer + the browser ONNX export, the card's pinned release revision 4aeb85b) \u2014 the tree a consumer downloads incl. onnx_browser/",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "hotchpotch/bekko-system-one-v0-400m",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "hotchpotch/bekko-system-one-v0-400m"
     }
    ]
   },
   "model_stack": [
    {
     "label": "bekko-400m model tree",
     "kind": "weights",
     "bytes": 3012643664,
     "src": 0
    }
   ],
   "targets": [
    "python env",
    "CPU (FP32 reference posture)"
   ],
   "note": "the bench board's bekko comparison lane \u2014 the seat moved 68M\u2192400M (reflex Bench 107, Plan 617 A6: the 400M beats the 68M on all 9 suites); MIT (verified 2026-10-02); subprocess oracle on loopback, measured by our harness"
  },
  {
   "key": "gliner",
   "name": "GLiNER2.5-Decide",
   "framework": "their gliner2 package in a torch-cu venv",
   "engine_kind": "python",
   "engine_bytes": 4440404787,
   "engine_what": "the gliner2 venv (torch-cu + transformers + peft + accelerate)",
   "engine_provenance": {
    "source": "recorded measurement",
    "detail": "the gliner2 venv (torch-cu + transformers + peft + accelerate + their gliner2 package) \u2014 measured 2026-09-25 on 4090-win: PowerShell (Get-ChildItem -Recurse -Force -File | Measure-Object -Sum Length).Sum over E:\\git\\riir-reflex\\.raw\\gliner-env \u2192 4234.7 MiB \u00d7 1048576",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "gliner2 venv",
      "what": "the gliner2 venv (torch-cu + transformers + peft + accelerate + their gliner2 package)",
      "date_utc": "2026-09-25",
      "host": "4090-win",
      "how": "PowerShell (Get-ChildItem -Recurse -Force -File | Measure-Object -Sum Length).Sum over E:\\git\\riir-reflex\\.raw\\gliner-env \u2192 4234.7 MiB \u00d7 1048576"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "gliner2 venv",
     "kind": "runtime",
     "bytes": 4440404787,
     "src": 0
    }
   ],
   "model_bytes": 1954845268,
   "model_what": "their model + tokenizer tree (HF, exact bytes)",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "fastino/GLiNER2.5-Decide",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "fastino/GLiNER2.5-Decide"
     }
    ]
   },
   "model_stack": [
    {
     "label": "model + tokenizer tree",
     "kind": "weights",
     "bytes": 1954845268,
     "src": 0
    }
   ],
   "targets": [
    "python env",
    "GPU (CUDA)"
   ],
   "note": null
  },
  {
   "key": "clef",
   "name": "clef-flash (9B) \u00b7 local 4-bit",
   "framework": "the community MLX port (clef_mlx.py) in an mlx venv, behind a loopback HTTP shim",
   "engine_kind": "python",
   "engine_bytes": 584762455,
   "engine_what": "the clef-flash (9B) lane venv (python 3.12: mlx + mlx-lm + mlx-vlm \u2014 the MLX port's tested pins)",
   "engine_provenance": {
    "source": "recorded measurement",
    "detail": "the clef-flash (9B) lane venv (python 3.12: mlx 0.32.3 + mlx-lm 0.32.0 + mlx-vlm 0.7.4 \u2014 the community MLX port's tested pins, riir-reflex Bench 113) \u2014 measured 2026-10-04 on m3-max-metal: uv venv --python 3.12 /tmp/clef-env + `uv pip install mlx==0.32.3 mlx-lm==0.32.0 mlx-vlm==0.7.4` (Bench 113's recorded pins; the bench-time env was an ephemeral uv run, so this rebuilds it from the pins); sum of file sizes (584,274,775 over 10,506 files) + dir entries (487,680), the lstat convention of the sibling venv rows.",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "mlx venv",
      "what": "the clef-flash (9B) lane venv (python 3.12: mlx 0.32.3 + mlx-lm 0.32.0 + mlx-vlm 0.7.4 \u2014 the community MLX port's tested pins, riir-reflex Bench 113)",
      "date_utc": "2026-10-04",
      "host": "m3-max-metal",
      "how": "uv venv --python 3.12 /tmp/clef-env + `uv pip install mlx==0.32.3 mlx-lm==0.32.0 mlx-vlm==0.7.4` (Bench 113's recorded pins; the bench-time env was an ephemeral uv run, so this rebuilds it from the pins); sum of file sizes (584,274,775 over 10,506 files) + dir entries (487,680), the lstat convention of the sibling venv rows."
     }
    ]
   },
   "engine_stack": [
    {
     "label": "mlx venv",
     "kind": "runtime",
     "bytes": 584762455,
     "src": 0
    }
   ],
   "model_bytes": 6213978733,
   "model_what": "their Clef-flash (9B) decision model as the community 4-bit MLX quant (two safetensors shards + the joint decision head + tokenizer) \u2014 the tree a consumer downloads",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "mlx-community/clef-flash-4bit",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "mlx-community/clef-flash-4bit"
     }
    ]
   },
   "model_stack": [
    {
     "label": "clef-flash (9B) 4-bit tree",
     "kind": "weights",
     "bytes": 6213978733,
     "src": 0
    }
   ],
   "targets": [
    "python env",
    "Apple GPU (MLX)"
   ],
   "note": "the bench board's clef-flash (9B) local comparison lane (Cloudflare/clef-flash, Apache-2.0, not affiliated; their larger Clef (27B) is not measured) \u2014 the LOCAL posture the board measures; the hosted Workers-AI posture is owner-gated and never pooled with it"
  },
  {
   "key": "agentjev",
   "name": "AgentJev-0.6B",
   "framework": "their jev_service in a torch venv",
   "engine_kind": "python",
   "engine_bytes": 4518943130,
   "engine_what": "their service venv (torch-cu + deps)",
   "engine_provenance": {
    "source": "recorded measurement",
    "detail": "their jev_service venv (torch-cu + deps) \u2014 measured 2026-09-25 on 4090-win: PowerShell (Get-ChildItem -Recurse -Force -File | Measure-Object -Sum Length).Sum over E:\\git\\riir-reflex\\.raw\\agentjev-env \u2192 4309.6 MiB \u00d7 1048576",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "jev_service venv",
      "what": "their jev_service venv (torch-cu + deps)",
      "date_utc": "2026-09-25",
      "host": "4090-win",
      "how": "PowerShell (Get-ChildItem -Recurse -Force -File | Measure-Object -Sum Length).Sum over E:\\git\\riir-reflex\\.raw\\agentjev-env \u2192 4309.6 MiB \u00d7 1048576"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "jev_service venv",
     "kind": "runtime",
     "bytes": 4518943130,
     "src": 0
    }
   ],
   "model_bytes": 3925763870,
   "model_what": "their agent-jev tensors + the Qwen3-0.6B backbone their service loads at boot",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "aimeigaoshou/agent-jev + Qwen/Qwen3-0.6B",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "aimeigaoshou/agent-jev + Qwen/Qwen3-0.6B"
     }
    ]
   },
   "model_stack": [
    {
     "label": "agent-jev + Qwen3-0.6B",
     "kind": "weights",
     "bytes": 3925763870,
     "src": 0
    }
   ],
   "targets": [
    "python env",
    "GPU (CUDA)"
   ],
   "note": "their boot loads BOTH the agent-jev checkpoint and the Qwen3-0.6B base from the HF cache \u2014 both counted"
  },
  {
   "key": "pplx",
   "name": "pplx-decider (27B) \u00b7 local 4-bit",
   "framework": "the MLX 4-bit port behind a loopback adapter (the reference wire)",
   "engine_kind": "python",
   "engine_bytes": 609572881,
   "engine_what": "the pplx lane venv (python 3.12: mlx + mlx-vlm + transformers \u2014 the port's tested pins)",
   "engine_provenance": {
    "source": "recorded measurement",
    "detail": "the pplx lane venv (python 3.12: mlx 0.32.3 + mlx-vlm 0.7.6 + transformers 5.19.0 \u2014 the MLX port's tested pins, riir-reflex Bench 132) + the adapter rig \u2014 measured 2026-10-09 on m3-max-metal: the bench-time env itself (.raw/pplx_srv/.venv, uv-built from the port's requirements.txt): sum of file sizes (609,047,889 over 11,246 files) + dir entries (524,992), the lstat convention of the sibling venv rows",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "mlx venv",
      "what": "the pplx lane venv (python 3.12: mlx 0.32.3 + mlx-vlm 0.7.6 + transformers 5.19.0 \u2014 the MLX port's tested pins, riir-reflex Bench 132) + the adapter rig",
      "date_utc": "2026-10-09",
      "host": "m3-max-metal",
      "how": "the bench-time env itself (.raw/pplx_srv/.venv, uv-built from the port's requirements.txt): sum of file sizes (609,047,889 over 11,246 files) + dir entries (524,992), the lstat convention of the sibling venv rows"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "mlx venv",
     "kind": "runtime",
     "bytes": 609572881,
     "src": 0
    }
   ],
   "model_bytes": 15362258534,
   "model_what": "Perplexity's pplx-decider-v1.1-27b as the community 4-bit MLX quant (affine 4-bit g64 backbone; readout + vision tower bf16) \u2014 the tree a consumer downloads",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "mlx-community/pplx-decider-v1.1-27b-4bit",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "mlx-community/pplx-decider-v1.1-27b-4bit"
     }
    ]
   },
   "model_stack": [
    {
     "label": "pplx-decider (27B) 4-bit tree",
     "kind": "weights",
     "bytes": 15362258534,
     "src": 0
    }
   ],
   "targets": [
    "python env",
    "Apple GPU (MLX)"
   ],
   "note": "the bench board's pplx comparison lane (perplexity-ai/pplx-decider-v1.1-27b, Apache-2.0, not affiliated \u2014 the JDI v3 board's #1): the LOCAL posture the board measures, the clef lane's class exactly; upstream bf16 weights are not downloaded"
  },
  {
   "key": "drex",
   "name": "Drex-DLM (8B)",
   "framework": "their serve.py in a torch-cu venv behind the TypeSafe-wire HTTP shim",
   "engine_kind": "python",
   "engine_bytes": 4871417003,
   "engine_what": "their service venv (torch cu124 + transformers \u2014 their serve.py's tested pins)",
   "engine_provenance": {
    "source": "recorded measurement",
    "detail": "the drex lane venv (.raw/drex-env: torch cu124 + transformers 5.19 \u2014 their serve.py's tested pins, riir-reflex Issue 073/Bench 125) \u2014 measured 2026-10-12 on 4090-win: the bench-time env itself (.raw/drex-env on the 4090 box): sum of file sizes (4,871,415,165 over 20,623 files) + dir entries (1,838), the lstat convention of the sibling venv rows",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "torch-cu venv",
      "what": "the drex lane venv (.raw/drex-env: torch cu124 + transformers 5.19 \u2014 their serve.py's tested pins, riir-reflex Issue 073/Bench 125)",
      "date_utc": "2026-10-12",
      "host": "4090-win",
      "how": "the bench-time env itself (.raw/drex-env on the 4090 box): sum of file sizes (4,871,415,165 over 20,623 files) + dir entries (1,838), the lstat convention of the sibling venv rows"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "torch-cu venv",
     "kind": "runtime",
     "bytes": 4871417003,
     "src": 0
    }
   ],
   "model_bytes": 16403005719,
   "model_what": "the Drex-DLM decision model as the full BF16 safetensors tree (4 shards + the decision head.pt + tokenizer) \u2014 the tree a consumer downloads",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "nace-ai/drex-dlm",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "nace-ai/drex-dlm"
     }
    ]
   },
   "model_stack": [
    {
     "label": "drex-dlm (Efficient-DLM-8B) BF16 tree",
     "kind": "weights",
     "bytes": 16403005719,
     "src": 0
    }
   ],
   "targets": [
    "python env",
    "GPU (CUDA)"
   ],
   "note": "the bench board's drex comparison lane (Nace.AI drex-dlm @ 6c63df2, backbone nvidia/Efficient-DLM-8B): repo code MIT, model weights CC BY-NC 4.0 \u2014 measurement only: comparison lane yes, distill teacher no, product serving no. The board's second posture (their llama.cpp edlm fork, Q8_0 on Metal) reads the same decision model re-quantized \u2014 one download, two runtimes"
  },
  {
   "key": "clm",
   "name": "CLM v0.1-8B",
   "framework": "vLLM docker image + their clm-serve head",
   "engine_kind": "python",
   "engine_bytes": 30777804339,
   "engine_what": "the vLLM serving image + their CLM repo/head checkout",
   "engine_provenance": {
    "source": "recorded measurement (sum)",
    "detail": "30,700,000,000 B (the vLLM serving image (vllm/vllm-openai:nightly) their CLM stack boots in, 2026-09-25 on 4090-win) + 77,804,339 B (their CLM repo checkout @ cca045ff + the trained head (checkpoints/CLM_v0.1-8B.pt), 2026-09-25 on 4090-win)",
    "parts": [
     {
      "source": "recorded measurement",
      "label": "vLLM docker image",
      "what": "the vLLM serving image (vllm/vllm-openai:nightly) their CLM stack boots in",
      "date_utc": "2026-09-25",
      "host": "4090-win",
      "how": "docker images --format '{{.Repository}}:{{.Tag}} {{.Size}}' \u2192 '30.7GB' (docker-reported, one decimal, decimal GB)"
     },
     {
      "source": "recorded measurement",
      "label": "CLM repo + trained head",
      "what": "their CLM repo checkout @ cca045ff + the trained head (checkpoints/CLM_v0.1-8B.pt)",
      "date_utc": "2026-09-25",
      "host": "4090-win",
      "how": "PowerShell recursive file-length sum over E:\\git\\riir-reflex\\.raw\\CLM \u2192 74.2 MiB \u00d7 1048576"
     }
    ]
   },
   "engine_stack": [
    {
     "label": "vLLM docker image",
     "kind": "runtime",
     "bytes": 30700000000,
     "src": 0
    },
    {
     "label": "CLM repo + trained head",
     "kind": "runtime",
     "bytes": 77804339,
     "src": 1
    }
   ],
   "model_bytes": 16473292630,
   "model_what": "the Qwen3-8B encoder weights + the trained head (CLM_v0.1-8B.pt)",
   "model_provenance": {
    "source": "huggingface.co tree API (exact bytes)",
    "detail": "Qwen/Qwen3-8B + Contrastive-LM/CLM-v0.1-8B",
    "parts": [
     {
      "source": "huggingface.co tree API (exact bytes)",
      "what": "Qwen/Qwen3-8B + Contrastive-LM/CLM-v0.1-8B"
     }
    ]
   },
   "model_stack": [
    {
     "label": "Qwen3-8B + CLM head",
     "kind": "weights",
     "bytes": 16473292630,
     "src": 0
    }
   ],
   "targets": [
    "docker container",
    "GPU (CUDA)"
   ],
   "note": null
  }
 ]
}