{
  "schema": "s2s-bench-v1",
  "kind": "data-kit-index",
  "exhibit": "flash-next-why-the-first-word-comes-sooner",
  "exhibit_title": "Why the first word comes sooner: the Strata engine and llama.cpp on Qwen3.8-Flash-Next, the full bench",
  "exhibit_url": "https://research.strata2signal.com/flash-next-why-the-first-word-comes-sooner/",
  "published_utc": "2026-10-13",
  "licence": "CC BY 4.0",
  "attribution": "strata→signal research, research.strata2signal.com",
  "contact": "hello@strata2signal.com",
  "run_id": "flash-next-strata-and-llama-cpp-2026-10-02",
  "run_window_utc": {
    "laptop_opened_utc": "2026-10-02T20:53:28Z",
    "laptop_closed_utc": "2026-10-05T21:20:47Z",
    "laptop_refused_start_utc": "2026-10-09T12:04:50Z",
    "desktop_96gb_llama_cpp_opened_utc": "2026-10-06T07:38:08Z",
    "desktop_96gb_llama_cpp_closed_utc": "2026-10-06T07:45:02Z",
    "desktop_96gb_strata_opened_utc": "2026-10-07T07:38:21Z",
    "desktop_96gb_strata_closed_utc": "2026-10-07T07:40:57Z",
    "desktop_24gb_window_opened_utc": "2026-10-09T09:31:47Z",
    "desktop_24gb_window_closed_utc": "2026-10-09T10:51:44Z",
    "note": "Each attempt's own first and last status line, in UTC; the laptop's span runs from its first speed attempt (L-IQ3 a1) to its last (S-NEW a2), with the refused third S-NEW attempt beside it, and the 24 GB card's window from its first one-second line to its last."
  },
  "card_class": "an RTX 5090 Laptop GPU (24 GB) in a laptop with 64 GB of memory; an RTX PRO 6000 Blackwell (96 GB) at a 300 W cap; and an EVGA GeForce RTX 3090 Ti FTW3 Ultra (24 GB) at a 300 W cap, in the same desktop as the 96 GB card",
  "sanitisation": "Machine readers start at `index.json`, which lists every file in this kit but three, each with its byte count and its sha256; the three it leaves out are itself, `provenance.json` and this directory's index page. `provenance.json` is the other half: it names every rule applied to every file, with the count of times each one fired, and says whether any rule touched each file. `README.md` says what is here, how to read it and what is not here.",
  "files_note": "`index.json` lists every file in this kit but itself, `provenance.json` and this directory's index page, each with its sha256; `provenance.json` names every rule and how many times it fired on each file.",
  "file_count": 170,
  "files": [
    {
      "file": "README.md",
      "bytes": 18000,
      "sha256": "f0590077e783d91fbe8a2d56909489ab30a2469dadf638ecad9ecda428b46501",
      "role": "What this directory holds, how to read it, the build and launch lines as templates, and what is not here."
    },
    {
      "file": "SUMMARY.md",
      "bytes": 28747,
      "sha256": "89cdcc0559fadd8c429043a6c2dcd6a4bcd3b8702777ad4d12c1cfbbba1e670c",
      "role": "The laptop bench's summary as its harness generated it; its section 3 holds the quality rows as counts (the six-way and the doorman), which is all of the quality rows this kit carries."
    },
    {
      "file": "build/desktop-24g-llamacpp-3cf03257-strata-6f32ec07-sm86.sh",
      "bytes": 8004,
      "sha256": "fed318246c78d8f375a7c395afbae4c2765753323291739907cb99a917feeacf",
      "role": "The build scripts as they ran: llama.cpp and the Strata engine, each from source at the pinned commit, for each card's architecture."
    },
    {
      "file": "build/desktop-24g-llamacpp-b11514-sm86.sh",
      "bytes": 7436,
      "sha256": "135e5e6d9b0b0634ebbd1c742e214bc97053d72c0943272cbbdf0bf7c08651cb",
      "role": "The build scripts as they ran: llama.cpp and the Strata engine, each from source at the pinned commit, for each card's architecture."
    },
    {
      "file": "build/desktop-llamacpp-3cf03257-sm120.sh",
      "bytes": 2145,
      "sha256": "b07713ff571e936b8092e2ca96543fbfe17c32bddb543f220747fd49571edc99",
      "role": "The build scripts as they ran: llama.cpp and the Strata engine, each from source at the pinned commit, for each card's architecture."
    },
    {
      "file": "build/desktop-strata-6f32ec07-sm120.sh",
      "bytes": 5241,
      "sha256": "730e966bcb4b00d1a7e80105efbef70726adca0799b656617963653df316fa92",
      "role": "The build scripts as they ran: llama.cpp and the Strata engine, each from source at the pinned commit, for each card's architecture."
    },
    {
      "file": "build/laptop-llamacpp-3cf03257-sm120.sh",
      "bytes": 952,
      "sha256": "62d0f142d9557115d4b192b39a543cacfc897bf8c465385106be270710359fbf",
      "role": "The build scripts as they ran: llama.cpp and the Strata engine, each from source at the pinned commit, for each card's architecture."
    },
    {
      "file": "build/laptop-strata-6f32ec07-sm120.sh",
      "bytes": 2615,
      "sha256": "a74bf3f679833bb4943cf4cf22b939853d5ba51119d8da82f633a3008d3d089e",
      "role": "The build scripts as they ran: llama.cpp and the Strata engine, each from source at the pinned commit, for each card's architecture."
    },
    {
      "file": "build/laptop-strata-8fc40dde-36fa455e-sm120.sh",
      "bytes": 1347,
      "sha256": "0495fddef940c9319b41a29c524928f3ed603d81b63d19ce92905c7284400f67",
      "role": "The build scripts as they ran: llama.cpp and the Strata engine, each from source at the pinned commit, for each card's architecture."
    },
    {
      "file": "client/common.py",
      "bytes": 42068,
      "sha256": "4f65a6e809a10061a828fdd65ed65179ca651cd2c273e9eaac14285fd19404ae",
      "role": "The client that sent every request and timed it: the step order inside the sandbox (`inner.py`), the engines' start-up and log readers, the prompt builder, the shared constants, the sandbox units, and from the house's earlier harness the stream reader with the timing rule of record (`vendor/vllm-seats/engine.py`), llama.cpp's device gate, the offload gate and the seeded ticket. All three desktop windows ran these bytes; the laptop ran the same `inner.py`, `engines.py`, `prompts.py`, `units.py` and vendored files. They are a reading set, not a program you can run: they import modules this kit does not carry (`quality.py`, `egress.py` and the rest of the earlier harness)."
    },
    {
      "file": "client/engines.py",
      "bytes": 40068,
      "sha256": "78110bba4685b2da5ba691e8e3099f959fe4fc928cf1e779527773235edae939",
      "role": "The client that sent every request and timed it: the step order inside the sandbox (`inner.py`), the engines' start-up and log readers, the prompt builder, the shared constants, the sandbox units, and from the house's earlier harness the stream reader with the timing rule of record (`vendor/vllm-seats/engine.py`), llama.cpp's device gate, the offload gate and the seeded ticket. All three desktop windows ran these bytes; the laptop ran the same `inner.py`, `engines.py`, `prompts.py`, `units.py` and vendored files. They are a reading set, not a program you can run: they import modules this kit does not carry (`quality.py`, `egress.py` and the rest of the earlier harness)."
    },
    {
      "file": "client/inner.py",
      "bytes": 28195,
      "sha256": "4980724e8fd8a911912d06b58b2fa74c6feb9a8fd126da6a05377b66e2d1a4f3",
      "role": "The client that sent every request and timed it: the step order inside the sandbox (`inner.py`), the engines' start-up and log readers, the prompt builder, the shared constants, the sandbox units, and from the house's earlier harness the stream reader with the timing rule of record (`vendor/vllm-seats/engine.py`), llama.cpp's device gate, the offload gate and the seeded ticket. All three desktop windows ran these bytes; the laptop ran the same `inner.py`, `engines.py`, `prompts.py`, `units.py` and vendored files. They are a reading set, not a program you can run: they import modules this kit does not carry (`quality.py`, `egress.py` and the rest of the earlier harness)."
    },
    {
      "file": "client/prompts.py",
      "bytes": 3933,
      "sha256": "500aeb228f7475cb04f249799d8510ab9fa3e69eba8470c8deea7cb4ea42e50a",
      "role": "The client that sent every request and timed it: the step order inside the sandbox (`inner.py`), the engines' start-up and log readers, the prompt builder, the shared constants, the sandbox units, and from the house's earlier harness the stream reader with the timing rule of record (`vendor/vllm-seats/engine.py`), llama.cpp's device gate, the offload gate and the seeded ticket. All three desktop windows ran these bytes; the laptop ran the same `inner.py`, `engines.py`, `prompts.py`, `units.py` and vendored files. They are a reading set, not a program you can run: they import modules this kit does not carry (`quality.py`, `egress.py` and the rest of the earlier harness)."
    },
    {
      "file": "client/units.py",
      "bytes": 8162,
      "sha256": "d74c207cf5bbea55a96eec73f114f740f9a3219907c0936307fc922bd30ef05f",
      "role": "The client that sent every request and timed it: the step order inside the sandbox (`inner.py`), the engines' start-up and log readers, the prompt builder, the shared constants, the sandbox units, and from the house's earlier harness the stream reader with the timing rule of record (`vendor/vllm-seats/engine.py`), llama.cpp's device gate, the offload gate and the seeded ticket. All three desktop windows ran these bytes; the laptop ran the same `inner.py`, `engines.py`, `prompts.py`, `units.py` and vendored files. They are a reading set, not a program you can run: they import modules this kit does not carry (`quality.py`, `egress.py` and the rest of the earlier harness)."
    },
    {
      "file": "client/vendor/vllm-seats/engine.py",
      "bytes": 21435,
      "sha256": "4ad930e6677c2996e385849d9f4c4e4d19aaad34069490f9266b2badf6427e5e",
      "role": "The client that sent every request and timed it: the step order inside the sandbox (`inner.py`), the engines' start-up and log readers, the prompt builder, the shared constants, the sandbox units, and from the house's earlier harness the stream reader with the timing rule of record (`vendor/vllm-seats/engine.py`), llama.cpp's device gate, the offload gate and the seeded ticket. All three desktop windows ran these bytes; the laptop ran the same `inner.py`, `engines.py`, `prompts.py`, `units.py` and vendored files. They are a reading set, not a program you can run: they import modules this kit does not carry (`quality.py`, `egress.py` and the rest of the earlier harness)."
    },
    {
      "file": "client/vendor/vllm-seats/engine_llamaserver.py",
      "bytes": 11881,
      "sha256": "77f1bcfc6b7f65df6396bd0922a27bb85bde76fd7ebec82b5e9f88c3b7ceed9e",
      "role": "The client that sent every request and timed it: the step order inside the sandbox (`inner.py`), the engines' start-up and log readers, the prompt builder, the shared constants, the sandbox units, and from the house's earlier harness the stream reader with the timing rule of record (`vendor/vllm-seats/engine.py`), llama.cpp's device gate, the offload gate and the seeded ticket. All three desktop windows ran these bytes; the laptop ran the same `inner.py`, `engines.py`, `prompts.py`, `units.py` and vendored files. They are a reading set, not a program you can run: they import modules this kit does not carry (`quality.py`, `egress.py` and the rest of the earlier harness)."
    },
    {
      "file": "client/vendor/vllm-seats/gates.py",
      "bytes": 13658,
      "sha256": "4565956a74c76470047a5b49091ad5bfdc4d3be492bd48154b1ddb2f5e3a9d05",
      "role": "The client that sent every request and timed it: the step order inside the sandbox (`inner.py`), the engines' start-up and log readers, the prompt builder, the shared constants, the sandbox units, and from the house's earlier harness the stream reader with the timing rule of record (`vendor/vllm-seats/engine.py`), llama.cpp's device gate, the offload gate and the seeded ticket. All three desktop windows ran these bytes; the laptop ran the same `inner.py`, `engines.py`, `prompts.py`, `units.py` and vendored files. They are a reading set, not a program you can run: they import modules this kit does not carry (`quality.py`, `egress.py` and the rest of the earlier harness)."
    },
    {
      "file": "client/vendor/vllm-seats/nonces.py",
      "bytes": 5878,
      "sha256": "8e0e57b6cc8c995f9683692435f5303e08cf29b326edecc421c8fc3bb0964dc2",
      "role": "The client that sent every request and timed it: the step order inside the sandbox (`inner.py`), the engines' start-up and log readers, the prompt builder, the shared constants, the sandbox units, and from the house's earlier harness the stream reader with the timing rule of record (`vendor/vllm-seats/engine.py`), llama.cpp's device gate, the offload gate and the seeded ticket. All three desktop windows ran these bytes; the laptop ran the same `inner.py`, `engines.py`, `prompts.py`, `units.py` and vendored files. They are a reading set, not a program you can run: they import modules this kit does not carry (`quality.py`, `egress.py` and the rest of the earlier harness)."
    },
    {
      "file": "records/L-IQ3-t8/speed/a1/RESULT.json",
      "bytes": 12719,
      "sha256": "09561d726d28a4e3235d4f576454afa91b0dba18238962571298370b0df7d841",
      "role": "The attempt's run record as the driver wrote it when the attempt ran to its end: the engine line, the pinned files and their digests, the operating system and driver, the memory plan, the sandbox and its network probe, and how the attempt ended."
    },
    {
      "file": "records/L-IQ3-t8/speed/a1/rows.jsonl",
      "bytes": 37038,
      "sha256": "4c05e7ce970fd51784d5dd44d19a939fe25f74c72200009a9389ef2dc87fbb15",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/L-IQ3-t8/speed/a1/spec.json",
      "bytes": 1248,
      "sha256": "327e95678b93b46314039927a59f11189c1b0af739aae36e1f309bf1599330bc",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/L-IQ3-t8/speed/a1/status.jsonl",
      "bytes": 5208,
      "sha256": "1d904100e6f7b0cac995b1e10c9cdf2c0d710d9b04abd01149508eb409749641",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/L-IQ3-t8/speed/a1/telemetry.csv",
      "bytes": 377411,
      "sha256": "edceab1d4a28843d4d23bd4917591634db415feb9e1a7a74ccb843dbf27bfc2b",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/L-IQ3-ub/speed/a1/RESULT.json",
      "bytes": 14952,
      "sha256": "883e3a49596ea5bdb3357ed1a9cbf3493faa61b3cfa3d6b9a96aa45ae881774e",
      "role": "The attempt's run record as the driver wrote it when the attempt ran to its end: the engine line, the pinned files and their digests, the operating system and driver, the memory plan, the sandbox and its network probe, and how the attempt ended."
    },
    {
      "file": "records/L-IQ3-ub/speed/a1/rows.jsonl",
      "bytes": 37290,
      "sha256": "1046cb90587dfa8285e37a8df514894f07a9add8888f5fcd1a858a0313e69a77",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/L-IQ3-ub/speed/a1/spec.json",
      "bytes": 1285,
      "sha256": "cea407504fdd2852a24f2a8fe707e531d296fc62f7f0b090a40b841485ea54e7",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/L-IQ3-ub/speed/a1/status.jsonl",
      "bytes": 5208,
      "sha256": "7274040e191734b9fcb704188dba68e5fccc192a51409fbda10467a88bb4fc4c",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/L-IQ3-ub/speed/a1/telemetry.csv",
      "bytes": 160384,
      "sha256": "a9d287755ec29cebdac209da2748cee870706591f7284b621413bc0a5d269a32",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/L-IQ3/speed/a1/RESULT.json",
      "bytes": 9671,
      "sha256": "a342aae624514d7a0fcc32feeb8df2f9c1b2f902dcb49cb17559721916eb61e3",
      "role": "The attempt's run record as the driver wrote it when the attempt ran to its end: the engine line, the pinned files and their digests, the operating system and driver, the memory plan, the sandbox and its network probe, and how the attempt ended."
    },
    {
      "file": "records/L-IQ3/speed/a1/rows.jsonl",
      "bytes": 28334,
      "sha256": "1f195569f2c607486a8cb3582a1ce66cafbb6e6c8cb73b85bea5d9e53ff4ee30",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/L-IQ3/speed/a1/spec.json",
      "bytes": 1149,
      "sha256": "a852576f52a6dd86000943acdb878e5d5778f833f137a81003ed098a735eba90",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/L-IQ3/speed/a1/status.jsonl",
      "bytes": 4496,
      "sha256": "a4a9aa83009ef78482d7de1f42889ded141943bb71b2698419cfbde0d765c017",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/L-IQ3/speed/a1/telemetry.csv",
      "bytes": 342838,
      "sha256": "306771293f3030a1c3c251548e3a3706b53ed504abd339d8b0cdeb4e93f12653",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/L-Q2/speed/a1/RESULT.json",
      "bytes": 14520,
      "sha256": "9451edbc04ee1abae18b446115532e8bab1cd80d228609acd82b613b5f143d98",
      "role": "The attempt's run record as the driver wrote it when the attempt ran to its end: the engine line, the pinned files and their digests, the operating system and driver, the memory plan, the sandbox and its network probe, and how the attempt ended."
    },
    {
      "file": "records/L-Q2/speed/a1/rows.jsonl",
      "bytes": 36879,
      "sha256": "0d700efe4c2de1ed5d3f11095458d25fc10a61294f2ef782c423c8663f6f1f5c",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/L-Q2/speed/a1/spec.json",
      "bytes": 1236,
      "sha256": "ef3547e0ef0fa4e0dbadc6f5ebeb03a833e757d6381e3522169dbfc0e64c12bf",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/L-Q2/speed/a1/status.jsonl",
      "bytes": 5199,
      "sha256": "51cb920a2b87763fe1915d86a419a2114d3ac703755771f1cfb6bef3d618a45a",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/L-Q2/speed/a1/telemetry.csv",
      "bytes": 223276,
      "sha256": "6723147c821848de45d2e521f6f96f7bb14abff5317cf246ee682c9868d3beb5",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/L-Q2/speed/a2/RESULT.json",
      "bytes": 15966,
      "sha256": "0344053a2245e4e14638ff78e76604c011dd058504eeb385ca3b8e76d3063bc1",
      "role": "The attempt's run record as the driver wrote it when the attempt ran to its end: the engine line, the pinned files and their digests, the operating system and driver, the memory plan, the sandbox and its network probe, and how the attempt ended."
    },
    {
      "file": "records/L-Q2/speed/a2/rows.jsonl",
      "bytes": 36890,
      "sha256": "8c2024bb22630b000f739e57ab4b62b15a1f4069dd2f020affc9ef905dfa638c",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/L-Q2/speed/a2/spec.json",
      "bytes": 1236,
      "sha256": "1877c70d4b638023df4752691682a0fb8a6f63a6d9a5490b8fd7d2b4ff7c6442",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/L-Q2/speed/a2/status.jsonl",
      "bytes": 5200,
      "sha256": "5e9d91a0ae349fad78686cc66e39cc48878c37999407965db719c90041264413",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/L-Q2/speed/a2/telemetry.csv",
      "bytes": 239278,
      "sha256": "30b2c4409b380dc473b720058a92022820a825d41a3f7b44ed3e52cd58d423e4",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/arch.json",
      "bytes": 238,
      "sha256": "0a401e968d9876879be4569a593f8dd5e3d08147eb5a1253096fd90d98d843ba",
      "role": "The llama.cpp build's CUDA architecture, read from the server's start line before any request was measured."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/draft-head.json",
      "bytes": 695,
      "sha256": "f9d1f80d5192d58f473817bce06e5674685b799afcc4aed29a575a2e204ad71e",
      "role": "The check, before any request was measured, that the current llama.cpp build ran with no draft head, as registered (the model file carries no draft-head block): the server's own line that no speculative implementation was set."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/fences.json",
      "bytes": 843,
      "sha256": "615a1a5edd3dc00a9a3dc4a681b1bbd0369c0d8888dac7c46689be41cb3eb3fc",
      "role": "The two network fences around the engine and how each was proven to hold."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/gpu-1hz.csv",
      "bytes": 408856,
      "sha256": "d4bbf5ede262d7c98ef60a24faf456ddf329d2bfe33ba8f41860f9edd67a6c0c",
      "role": "The computer's one-second card log over the row, as nvidia-smi printed it (its clock is UTC, written without a zone), both cards by index, with each card's bus address withheld."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/identity.json",
      "bytes": 1927,
      "sha256": "d6fb2a7c9d361ed62b7bfa9e975537e1f58b3f2465207b41255ce476cd3a2004",
      "role": "The prompt identity check: each request's prompt digest against the laptop's."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/memory-receipts.jsonl",
      "bytes": 14854,
      "sha256": "1a4ed0fd0f84b75fbdd3f84caeb34235fbaa29f0e7a17b1e971f101e500125f9",
      "role": "The row's memory receipts: the engine unit's memory reading at each event, with the computer's free memory beside it."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/placement.json",
      "bytes": 915,
      "sha256": "6226128c522f79536c3d4e374b93276d3431f8d9fab8ba3b9c1a8c2a148e25f0",
      "role": "Where llama.cpp put the model: one CUDA device, the layers offloaded and the experts kept in system memory."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/row-meta.txt",
      "bytes": 4586,
      "sha256": "6238e85dedfa04260adac87689165630596f13615e003973b1676282d2dbe5b7",
      "role": "The row server's own account of the row: the pins it re-read, the engine line, the sandbox, the cards before and after, and how it ended."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/rows.jsonl",
      "bytes": 37286,
      "sha256": "2a885266bd1110c9cb44eafb6519e3dd1ad30499cefc9d19b97d31f0cad85690",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/run.json",
      "bytes": 5101,
      "sha256": "0cbb1e6def887179abd03c9e39fe1297f7c0f808f47d4473355211f37e95f535",
      "role": "The row's record as the window's row server wrote it: the engine's build identity, the fences that held, the residency or placement reading and how the row ended."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/spec.json",
      "bytes": 1237,
      "sha256": "aebba2f64278739cf633f3350ade8fd7e4aede5c9c58deec87be2d1da6bec775",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/R-LS-B11514-IQ3-24G/speed/a1/status.jsonl",
      "bytes": 4894,
      "sha256": "92a062b2e76d02d41f9b3663ec93edf169e2386edd7226ac1a3e210fa3156b92",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/arch.json",
      "bytes": 238,
      "sha256": "b13083e93d17dacd6673364a5f7d226bae882e0d56abb01bcb9e9e8cfb1bb377",
      "role": "The llama.cpp build's CUDA architecture, read from the server's start line before any request was measured."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/fences.json",
      "bytes": 836,
      "sha256": "4900d4fa59a64b115d884c83c5ca5cb5bf6678c958eb2c50222826e2be6c667e",
      "role": "The two network fences around the engine and how each was proven to hold."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/gpu-1hz.csv",
      "bytes": 448635,
      "sha256": "e5ea610b01f5ece305263e1fa6e14efc443c9a1d8fa0ab7198ad78cccf7a3474",
      "role": "The computer's one-second card log over the row, as nvidia-smi printed it (its clock is UTC, written without a zone), both cards by index, with each card's bus address withheld."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/identity.json",
      "bytes": 1927,
      "sha256": "d6fb2a7c9d361ed62b7bfa9e975537e1f58b3f2465207b41255ce476cd3a2004",
      "role": "The prompt identity check: each request's prompt digest against the laptop's."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/memory-receipts.jsonl",
      "bytes": 15140,
      "sha256": "de33e699021f122ffd7397707cae7adb11c55d87bb892235545946ed15f14b89",
      "role": "The row's memory receipts: the engine unit's memory reading at each event, with the computer's free memory beside it."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/placement.json",
      "bytes": 915,
      "sha256": "6226128c522f79536c3d4e374b93276d3431f8d9fab8ba3b9c1a8c2a148e25f0",
      "role": "Where llama.cpp put the model: one CUDA device, the layers offloaded and the experts kept in system memory."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/row-meta.txt",
      "bytes": 4248,
      "sha256": "6461193be0a37c0460652279cd3826b535c64267f2ae4872416253ca9cb09e89",
      "role": "The row server's own account of the row: the pins it re-read, the engine line, the sandbox, the cards before and after, and how it ended."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/rows.jsonl",
      "bytes": 37187,
      "sha256": "199b62827dd3fd2f97a0c4ae22353b65d81378db98e5ef78fa86dd2a1c406003",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/run.json",
      "bytes": 4304,
      "sha256": "728bc7e8ddac593993738dff38e76e1f248993a4bad21dd305e56c3208b1297d",
      "role": "The row's record as the window's row server wrote it: the engine's build identity, the fences that held, the residency or placement reading and how the row ended."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/spec.json",
      "bytes": 1218,
      "sha256": "4647f1bfdcd885614dc24c1cece84a9f3f932d24f53aaf32a24666b1099ab437",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/R-LS-IQ3-24G/speed/a1/status.jsonl",
      "bytes": 4894,
      "sha256": "9b54c8f70232c1e7552d15ec281c6c23be45e5948468e114daca5b0bd976fbf2",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/fences.json",
      "bytes": 859,
      "sha256": "45087ac8f619b1f951e1e09e6ae7e9da4eee6acccf538c7c72776fb32f198610",
      "role": "The two network fences around the engine and how each was proven to hold."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/gpu-1hz.csv",
      "bytes": 79272,
      "sha256": "b9d05e8130ab4aa1dba2d1cb81f4459712abdf32bbc2246e9a83074d2f027198",
      "role": "The computer's one-second card log over the row, as nvidia-smi printed it (its clock is UTC, written without a zone), both cards by index, with each card's bus address withheld."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/identity.json",
      "bytes": 1927,
      "sha256": "d6fb2a7c9d361ed62b7bfa9e975537e1f58b3f2465207b41255ce476cd3a2004",
      "role": "The prompt identity check: each request's prompt digest against the laptop's."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/memory-receipts.jsonl",
      "bytes": 12228,
      "sha256": "00084f9749a3a9b262d34e048febc8d08c8e471ccdacb07790341081640c9095",
      "role": "The row's memory receipts: the engine unit's memory reading at each event, with the computer's free memory beside it."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/placement.json",
      "bytes": 1163,
      "sha256": "3e214aaa4489938060fccdb3a1f4c610a755721366428dfac8dda9e13701871d",
      "role": "Where llama.cpp put the model: one CUDA device, the layers offloaded and the experts kept in system memory."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/row-meta.txt",
      "bytes": 3402,
      "sha256": "4a7fff73a0ed299d720e4c3b2c0f576c48e89fec08268b588e88f6061f981b92",
      "role": "The row server's own account of the row: the pins it re-read, the engine line, the sandbox, the cards before and after, and how it ended."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/rows.jsonl",
      "bytes": 37046,
      "sha256": "76f4475d25f71a159f7b047aedb34c587b07bb2206d4fa99e8a329746934f4b2",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/run.json",
      "bytes": 3452,
      "sha256": "bfe9215bd3c9796cc4cc0c334e081dd711099929e60a35aea77cbd477e316cff",
      "role": "The row's record as the window's row server wrote it: the engine's build identity, the fences that held, the residency or placement reading and how the row ended."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/spec.json",
      "bytes": 1191,
      "sha256": "4a0623366bd5896e0aaac1173c882672a7de470e93e24639514bf1a6724126d5",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/R-LS-IQ3/speed/a1/status.jsonl",
      "bytes": 4716,
      "sha256": "13ab8915e5d8782e402dd2a0b9a9a105bff84aabfd3eb3d7312b4ad22b69d0a7",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/fences.json",
      "bytes": 838,
      "sha256": "4086397775942eacb208dadd499878ec7df16ff0a8119b8e5133f850fe285690",
      "role": "The two network fences around the engine and how each was proven to hold."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/gpu-1hz.csv",
      "bytes": 53797,
      "sha256": "b61aba5389044b2f956e8325a82a0ba7248fc227fd203fcc1b7a258092b98af6",
      "role": "The computer's one-second card log over the row, as nvidia-smi printed it (its clock is UTC, written without a zone), both cards by index, with each card's bus address withheld."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/identity.json",
      "bytes": 1927,
      "sha256": "d6fb2a7c9d361ed62b7bfa9e975537e1f58b3f2465207b41255ce476cd3a2004",
      "role": "The prompt identity check: each request's prompt digest against the laptop's."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/memory-receipts.jsonl",
      "bytes": 15293,
      "sha256": "b66d205f01ab2b2b24d5a458a0a31726f4e29a3ca472951f7cb283d331507651",
      "role": "The row's memory receipts: the engine unit's memory reading at each event, with the computer's free memory beside it."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/residency.json",
      "bytes": 3218,
      "sha256": "049261f29d543275d3f2572d2c66154e33bb869bf9519d70f5958fd8427cdbe8",
      "role": "Where the Strata engine put the experts: the expert cache's slots on the card, read at the end of the row."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/row-meta.txt",
      "bytes": 5590,
      "sha256": "660d1f665c5f692f109080e9180da3f8dddde39350a13996f1f17ad44b3e8e58",
      "role": "The row server's own account of the row: the pins it re-read, the engine line, the sandbox, the cards before and after, and how it ended."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/rows.jsonl",
      "bytes": 44935,
      "sha256": "3bd169ea1d9ad17f09d007c569cfbd01252fae613dd2fab15c41ff1ac5be3bb9",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/run.json",
      "bytes": 7866,
      "sha256": "9521f94250b437249c8f9ed4b0af928c13b06c0dba0c180390b24917d34e6045",
      "role": "The row's record as the window's row server wrote it: the engine's build identity, the fences that held, the residency or placement reading and how the row ended."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/spec.json",
      "bytes": 1085,
      "sha256": "794878a77a6ff19cc9946437fb0f73e1477bee7a13008f9e7df835a1682d2d23",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/status.jsonl",
      "bytes": 5552,
      "sha256": "bc1aa74dc8fdf9e672a576dd8045c606a82f79e3c6d2e54c09462b47d5a5e3d9",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/strata-config.json",
      "bytes": 1123,
      "sha256": "4e4e043e949011736a03ce119ffd8e3feb4356e7b4650af385395d7149907aa2",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/strata-engine.log",
      "bytes": 6793,
      "sha256": "88d4fc79affd2c5d5a7ad49413adbb6f74ad89bae1160a31ba040580ea8819bc",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/R-S139-IQ3-24G/speed/a1/strata-line.json",
      "bytes": 2683,
      "sha256": "0c14f68fc90d901ef45e1a03bba80b06cf416e266ab5632fc6924dccc91f09da",
      "role": "The Strata engine's line on the desktop: the laptop's line, pinned, beside the line the engine's setup would have written for that computer."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/fences.json",
      "bytes": 859,
      "sha256": "b868dfff0089e849cece91a5e3c7e386440e8ece132c79896735a9bd484e7110",
      "role": "The two network fences around the engine and how each was proven to hold."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/gpu-1hz.csv",
      "bytes": 31213,
      "sha256": "b5728d4971de16626d67023a7394b2014abc76cc22a4c2965eb8854fc6bccf70",
      "role": "The computer's one-second card log over the row, as nvidia-smi printed it (its clock is UTC, written without a zone), both cards by index, with each card's bus address withheld."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/identity.json",
      "bytes": 1927,
      "sha256": "d6fb2a7c9d361ed62b7bfa9e975537e1f58b3f2465207b41255ce476cd3a2004",
      "role": "The prompt identity check: each request's prompt digest against the laptop's."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/memory-receipts.jsonl",
      "bytes": 13048,
      "sha256": "303aab086e7dce819ba0900954204c03007b1b839b4b96dbcfab7e6768df1792",
      "role": "The row's memory receipts: the engine unit's memory reading at each event, with the computer's free memory beside it."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/residency.json",
      "bytes": 1673,
      "sha256": "602bb26f2abfe0f25fd33f46fe725270cd675104fe50d3933f7568548a5375a1",
      "role": "Where the Strata engine put the experts: the expert cache's slots on the card, read at the end of the row."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/row-meta.txt",
      "bytes": 4301,
      "sha256": "8206732b5ede6ec1dd7357dbac28e5051aa56a9e1405b85dd333da1751114026",
      "role": "The row server's own account of the row: the pins it re-read, the engine line, the sandbox, the cards before and after, and how it ended."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/rows.jsonl",
      "bytes": 44530,
      "sha256": "29bf97d4b82a5d8f7276bc49edbfb2af5ce154cbfa8e37d5028d8a51dd3af198",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/run.json",
      "bytes": 7202,
      "sha256": "657d9eba3c2a68cb524a1fc3612d9ccb08f4d0b17aef95de4d2a0fe1a15653b7",
      "role": "The row's record as the window's row server wrote it: the engine's build identity, the fences that held, the residency or placement reading and how the row ended."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/spec.json",
      "bytes": 1020,
      "sha256": "8f14eb3e6ece0a756c1edfb3f78322d58a71f2b8aa80e361913e704a2de5244c",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/status.jsonl",
      "bytes": 5493,
      "sha256": "2480fc7ffb657dbe0369cde697236c8dbb9edbdaa051da2e622811ac65edbf40",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/strata-config.json",
      "bytes": 1100,
      "sha256": "dd31118f869313af43539bab1f143f529a73efa62359223d177820a3672d7908",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/strata-engine.log",
      "bytes": 5441,
      "sha256": "468487218f6c5550deaf1cfc72a3df8a2ed8bb1d5514cbfa506402b3e75ef37b",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/R-S139-IQ3/speed/a1/strata-line.json",
      "bytes": 2660,
      "sha256": "7d700e149cdf1bb069ce7987acabe1b5f6861bdafaa3d2a3abafeef2124e2165",
      "role": "The Strata engine's line on the desktop: the laptop's line, pinned, beside the line the engine's setup would have written for that computer."
    },
    {
      "file": "records/S-HEAD/speed/a1/RESULT.json",
      "bytes": 13617,
      "sha256": "6b62ea675d713fd0c8172e322f4f15fb438a4dba084db4f2daf4711d20e900e6",
      "role": "The attempt's run record as the driver wrote it when the attempt ran to its end: the engine line, the pinned files and their digests, the operating system and driver, the memory plan, the sandbox and its network probe, and how the attempt ended."
    },
    {
      "file": "records/S-HEAD/speed/a1/rows.jsonl",
      "bytes": 43546,
      "sha256": "5e6bf6a3e680e67b08da5f7bd3c4a6453da6aad6cc94d6a77c6d618086c516e8",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/S-HEAD/speed/a1/spec.json",
      "bytes": 1145,
      "sha256": "171e7c7c04a34d296d2d76c33e548209839cf5fb5914791797f64bacce47c265",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/S-HEAD/speed/a1/status.jsonl",
      "bytes": 5620,
      "sha256": "cefc4c23e273a875b7fe3818dadfd93bc3c5bdc997fc533398c47fc6dc0d08d1",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/S-HEAD/speed/a1/strata-config.json",
      "bytes": 1029,
      "sha256": "582589ded8be4b952e4e4635797f4617afb1542aadbf818397d2973309e78438",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/S-HEAD/speed/a1/strata-engine.log",
      "bytes": 6040,
      "sha256": "f4e3ca0815670da06b409cc990b5aefe64e12f27ffe2c9518e5247c0831e2612",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/S-HEAD/speed/a1/telemetry.csv",
      "bytes": 39208,
      "sha256": "d4d0d98d003f059eeda56c1187939049ddf96ecd6a6f47ef4490cc92ad323f18",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/S-HEAD/speed/a2/STOPPED.json",
      "bytes": 14542,
      "sha256": "4a14aedbebe2e1b0363faa0e625a2537ffa3dd868192e373a6cc73659c2b90ba",
      "role": "The attempt's run record as the driver wrote it when one of our guards stopped the attempt: the same facts as a finished attempt's, the guard that fired and its reading."
    },
    {
      "file": "records/S-HEAD/speed/a2/rows.jsonl",
      "bytes": 19339,
      "sha256": "8269379a471d01e09ab862052e15aaf19f06b72b5bc6a107052b59cad0ad39ce",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/S-HEAD/speed/a2/spec.json",
      "bytes": 1145,
      "sha256": "0dc86f23d1362cb73479a26094ea460efaa2d00b0c7b301b34f0e802901acdeb",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/S-HEAD/speed/a2/status.jsonl",
      "bytes": 4223,
      "sha256": "49491788b95a0fc5f83962d0838c220d85528833751f71357626632da3ad3148",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/S-HEAD/speed/a2/strata-config.json",
      "bytes": 1029,
      "sha256": "3d505357f9bd0a706a3b91b7b23edd5c58a6f0b4b5acd7f10afbc7615efad869",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/S-HEAD/speed/a2/strata-engine.log",
      "bytes": 4410,
      "sha256": "17fc385f9ddd13fd7f487c5819f0895930aae32079bf9b2d13e675e51874d129",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/S-HEAD/speed/a2/telemetry.csv",
      "bytes": 12117,
      "sha256": "b58789a984822b81b1e7873817ce5ae3793e1fad89cb8dfba1d93582cabc9080",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/S-IQ3/speed/a1/RESULT.json",
      "bytes": 13367,
      "sha256": "976ff0e7ee7296b4ed511e4353f60ef5d625e0c47963c8a7d7641bebe2270b6c",
      "role": "The attempt's run record as the driver wrote it when the attempt ran to its end: the engine line, the pinned files and their digests, the operating system and driver, the memory plan, the sandbox and its network probe, and how the attempt ended."
    },
    {
      "file": "records/S-IQ3/speed/a1/rows.jsonl",
      "bytes": 43490,
      "sha256": "c327f6a7cd386a06f5c47ad139036a093743b78f801896b3e36847b9f9dac834",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/S-IQ3/speed/a1/spec.json",
      "bytes": 1141,
      "sha256": "29693c1dda78df9aa71d4eec6c8b93880264d732c055a304d98b59432bafb67d",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/S-IQ3/speed/a1/status.jsonl",
      "bytes": 5380,
      "sha256": "18be9c962c46a8364550e0d6482fa351cec6c8dc0ccc4f4d563d9d97f329face",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/S-IQ3/speed/a1/strata-config.json",
      "bytes": 1001,
      "sha256": "b72461713928de3504b1d33410d2ae92206f1c7fb30249b099f38d5006b2c2d4",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/S-IQ3/speed/a1/strata-engine.log",
      "bytes": 5439,
      "sha256": "34764a8efd26adc4d1c92707c7418004e623ed788d5139315ebabbd3e86aaaa0",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/S-IQ3/speed/a1/telemetry.csv",
      "bytes": 40949,
      "sha256": "4e8ebf5efdd30f469bde18cc7d0ae50dec9882c1b86a8b3cc4d8d4d0c901b3de",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/S-IQ3/speed/a2/STOPPED.json",
      "bytes": 14079,
      "sha256": "b904759beed0d51a13af8d915291f8570e08743c4894fc25ccd2906a8c90c077",
      "role": "The attempt's run record as the driver wrote it when one of our guards stopped the attempt: the same facts as a finished attempt's, the guard that fired and its reading."
    },
    {
      "file": "records/S-IQ3/speed/a2/rows.jsonl",
      "bytes": 33880,
      "sha256": "22ef7a20a65e65a280b31e9a45b173374e598e60692173a0703dd7e1860ab9d2",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/S-IQ3/speed/a2/spec.json",
      "bytes": 1141,
      "sha256": "dc8278fa15b3f748b0ab63237f416c2fb4c7a15765328cbe71ed25b73c9ffecf",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/S-IQ3/speed/a2/status.jsonl",
      "bytes": 4840,
      "sha256": "26eef9cd5549e5a5c048ad5583dd0e1401fb1bc45e23b6ffa9dbb544c069bf95",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/S-IQ3/speed/a2/strata-config.json",
      "bytes": 1001,
      "sha256": "1506e5b88251143d666062ac444b4bfb332e9d926e2c6d6e1a2cd4a170aeed78",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/S-IQ3/speed/a2/strata-engine.log",
      "bytes": 4870,
      "sha256": "dc1772634c507697745dd0f631154e523dc32163a2e3e5b1dafa95024b82a2b7",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/S-IQ3/speed/a2/telemetry.csv",
      "bytes": 25254,
      "sha256": "3984b9c4b0686514217ee17703b88d0d98f54b3cc8c05f43b57ca511f31d1b7f",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/S-NEW/speed/PREFLIGHT-REFUSALS.jsonl",
      "bytes": 562,
      "sha256": "ce52624abcc53613f175986d7edfd6c1cee9eb5e6ca7f0d92c41a5be7ff42014",
      "role": "The refused start of the laptop's 0.1.39 arm's third attempt (2026-10-09): the memory check's sentence and its arithmetic. Nothing ran, so nothing from it is quoted."
    },
    {
      "file": "records/S-NEW/speed/a1/STOPPED.json",
      "bytes": 14439,
      "sha256": "afc5591a0b93b46513d22f1265907d6af10bfbb814e47c00bcc0c293a5611027",
      "role": "The attempt's run record as the driver wrote it when one of our guards stopped the attempt: the same facts as a finished attempt's, the guard that fired and its reading."
    },
    {
      "file": "records/S-NEW/speed/a1/rows.jsonl",
      "bytes": 14970,
      "sha256": "12ae1c1534033440334af898d938e2d3ae7cb1356104f900e27509996e380bfe",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/S-NEW/speed/a1/spec.json",
      "bytes": 1141,
      "sha256": "19fb2ecdc2322a90782d20ee7cb6743ccbf7cd61f178459c207ed60e79fa64e7",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/S-NEW/speed/a1/status.jsonl",
      "bytes": 4195,
      "sha256": "9081cca94fa9cf46bf671b83c22543b1f51e3aa2c187f10735c8bd27014ec7cd",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/S-NEW/speed/a1/strata-config.json",
      "bytes": 1072,
      "sha256": "22d29fd3596903069bca711c1960b5a955cd5c86b0b18eed24c0c4104f368061",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/S-NEW/speed/a1/strata-engine.log",
      "bytes": 4535,
      "sha256": "ba8ad18993526dfcb23136dd329861003109e6def798df8596d9c4127e5a0125",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/S-NEW/speed/a1/telemetry.csv",
      "bytes": 9628,
      "sha256": "b3944bc66783bbc13c8cc512893b1cb6944618c4efc923a1464f30bbc3c88893",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/S-NEW/speed/a2/STOPPED.json",
      "bytes": 14433,
      "sha256": "185c8fc1b7df0e128ccfb78e62592724ccf20e37df59634999b47962852c8fd1",
      "role": "The attempt's run record as the driver wrote it when one of our guards stopped the attempt: the same facts as a finished attempt's, the guard that fired and its reading."
    },
    {
      "file": "records/S-NEW/speed/a2/rows.jsonl",
      "bytes": 29941,
      "sha256": "67965b5856b334eea3f47f65f96edcea00e84e38f4a08e965b6c7d7724505538",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/S-NEW/speed/a2/spec.json",
      "bytes": 1141,
      "sha256": "9abb5e5bef7c67004e2e57a94a6451496c5d50c9395bb06719fb8a93fe1e764a",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/S-NEW/speed/a2/status.jsonl",
      "bytes": 5043,
      "sha256": "1557ec5c00c8d24f4e5dbc7f4e8e7bdaa6d55c0caa12751cdd13036f30f17631",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/S-NEW/speed/a2/strata-config.json",
      "bytes": 1072,
      "sha256": "467f9562eb287ce7ea2715b37a936e078f6dd2b3903b677564821c6da91c9c54",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/S-NEW/speed/a2/strata-engine.log",
      "bytes": 5535,
      "sha256": "448cd73db75f9fb0713618e5c71580ec7f24f802e2bdfd4e91237bde8f6e8e69",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/S-NEW/speed/a2/telemetry.csv",
      "bytes": 18350,
      "sha256": "17e247c7f533de4a01f6dd2ff9c3ae0dd779ee257cad69ba5ec733541ebeab34",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/S-Q2/speed/a1/RESULT.json",
      "bytes": 13408,
      "sha256": "18ceda7c6fcb431e0ebb53b01a35c61dfa143c35432b43b1c699c0e53a37878d",
      "role": "The attempt's run record as the driver wrote it when the attempt ran to its end: the engine line, the pinned files and their digests, the operating system and driver, the memory plan, the sandbox and its network probe, and how the attempt ended."
    },
    {
      "file": "records/S-Q2/speed/a1/rows.jsonl",
      "bytes": 43511,
      "sha256": "a6b2b6a26bfdc725368dc3ab91b3be5a8d94123b7c5d9af9fb2d020f3c0d9926",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/S-Q2/speed/a1/spec.json",
      "bytes": 1135,
      "sha256": "b46634512388e213401c149032978c7d1629fcc659188b955d3322b9ebd1ad59",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/S-Q2/speed/a1/status.jsonl",
      "bytes": 5382,
      "sha256": "0b7bd2ebfe11f791a5c96f16224115816613cb47ffa899b74c9481d2252a86fc",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/S-Q2/speed/a1/strata-config.json",
      "bytes": 1023,
      "sha256": "989abd0a84c59d99a4932815e7d608a74a489d10bc6e256ed0bc3da395891cda",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/S-Q2/speed/a1/strata-engine.log",
      "bytes": 6294,
      "sha256": "371d0f542d152dc61d874940f8a07baa8fa558e8612a342a2d4a49fcb6233d35",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/S-Q2/speed/a1/telemetry.csv",
      "bytes": 37166,
      "sha256": "04a9249471a438f7722fc91789f66af27abed179c6259c1f5e4f61f39b734e09",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/S-Q2/speed/a2/RESULT.json",
      "bytes": 14895,
      "sha256": "b4bb528a30cb6444feb59172752c297b204a3d76e72ca4276861b260fdb99a2f",
      "role": "The attempt's run record as the driver wrote it when the attempt ran to its end: the engine line, the pinned files and their digests, the operating system and driver, the memory plan, the sandbox and its network probe, and how the attempt ended."
    },
    {
      "file": "records/S-Q2/speed/a2/rows.jsonl",
      "bytes": 43535,
      "sha256": "a85dd6d96ea99144fee955e7a8be6993b859c9c120dedcb4bd66cd2acd90a8b8",
      "role": "The attempt's speed rows, the harness's own output: one line per request, with its prompt digest, the first-word wait and the writing speed on our client's clock, the engine's own timings beside them, the count check, and the card, CPU-package, disk and memory readings over the request's window. The model's text is withheld from every line (`completion_head`); its digest is kept."
    },
    {
      "file": "records/S-Q2/speed/a2/spec.json",
      "bytes": 1135,
      "sha256": "600299c2bee947a481bb78d7d0216eebe92e4bc335603e30f911535261dc61e5",
      "role": "The attempt's specification as the driver handed it to the client: the server's command line, the contexts, the repeats and the timeouts."
    },
    {
      "file": "records/S-Q2/speed/a2/status.jsonl",
      "bytes": 5384,
      "sha256": "399fbb31f71a8ae603b7ed7f859a605b330759c532ae3fe5ddd40711efb70b0b",
      "role": "The client's status lines for the attempt, one per step: the network probe, the launch, readiness, the start-up checks, each request and the end."
    },
    {
      "file": "records/S-Q2/speed/a2/strata-config.json",
      "bytes": 1023,
      "sha256": "4241110b57deb188cb4784f65b96bd1181da28dc00f111239287a6164badefa8",
      "role": "The Strata engine's configuration file for the attempt, as the client wrote it and the engine read it."
    },
    {
      "file": "records/S-Q2/speed/a2/strata-engine.log",
      "bytes": 6422,
      "sha256": "672ebda21dc52a65fc94721718b60a1cc233d3f618e13aad4a42eee2508981da",
      "role": "The Strata engine's own log for the attempt: its load lines, its expert-cache and prompt-path sizing, and one serve line per request."
    },
    {
      "file": "records/S-Q2/speed/a2/telemetry.csv",
      "bytes": 36619,
      "sha256": "6c234c6ee9b18d0b7ba32ff30b96fea4fa10f909fa749f0f3e55318b6b111595",
      "role": "The attempt's one-second telemetry on the laptop: card power, temperature, memory and clocks, the CPU package's energy counter, disk reads, load and memory pressure."
    },
    {
      "file": "records/window-24G/telemetry-1hz.csv",
      "bytes": 823260,
      "sha256": "859cc080efa5499cd49ae7d41ee28cc951a92566cad3d3464b2d6472c4e2a586",
      "role": "The 24 GB card's window, one line a second: both cards' power, cap, memory, temperature and use, the computer's free memory, the disk's temperature, the battery backup's status, the running arm's CPU time and the load average."
    },
    {
      "file": "registration/AMENDMENT-24G.md",
      "bytes": 67028,
      "sha256": "1ad586ee9639319264d0e20a1e8877aa483bf341741491905b14f707089633cc",
      "role": "The registration, written before the first reading it governs, with each amendment dated and appended: the laptop's and its nine amendments, the desktop's bands, the two desktop rows', the exploratory desktop row's, and the 24 GB card's with its addenda."
    },
    {
      "file": "registration/PREREG-BANDS.md",
      "bytes": 3745,
      "sha256": "b87d29c8154a3a908cf782374e3ff677d338234557604fb96eac90c4e65d9dc9",
      "role": "The registration, written before the first reading it governs, with each amendment dated and appended: the laptop's and its nine amendments, the desktop's bands, the two desktop rows', the exploratory desktop row's, and the 24 GB card's with its addenda."
    },
    {
      "file": "registration/PREREG-W0.md",
      "bytes": 19898,
      "sha256": "377abe9c5c171754ba86aa3cd6c0da0f62c6a64a57031760206ae38771b67dd1",
      "role": "The registration, written before the first reading it governs, with each amendment dated and appended: the laptop's and its nine amendments, the desktop's bands, the two desktop rows', the exploratory desktop row's, and the 24 GB card's with its addenda."
    },
    {
      "file": "registration/PREREG-W1.md",
      "bytes": 20822,
      "sha256": "cc6e5b6bae32b121bb835fe8495a02f84bf39e492007659d4df10377ac2a4f21",
      "role": "The registration, written before the first reading it governs, with each amendment dated and appended: the laptop's and its nine amendments, the desktop's bands, the two desktop rows', the exploratory desktop row's, and the 24 GB card's with its addenda."
    },
    {
      "file": "registration/PREREG-W2.md",
      "bytes": 33806,
      "sha256": "487216620e4e9b062d25907c1c9bb1bc30bc4a4c98ff221d243a924663ec6251",
      "role": "The registration, written before the first reading it governs, with each amendment dated and appended: the laptop's and its nine amendments, the desktop's bands, the two desktop rows', the exploratory desktop row's, and the 24 GB card's with its addenda."
    },
    {
      "file": "registration/PREREG.md",
      "bytes": 96722,
      "sha256": "7d3d25f91ec10423d9d4bdeefb6c6d6a6486fed9e8e3a1fee04bfa20fb426db9",
      "role": "The registration, written before the first reading it governs, with each amendment dated and appended: the laptop's and its nine amendments, the desktop's bands, the two desktop rows', the exploratory desktop row's, and the 24 GB card's with its addenda."
    },
    {
      "file": "requests/SLICES.json",
      "bytes": 591724,
      "sha256": "c8348533e2e90abe2f2f4c9ef9cee6f047b047972b423d18dacc4fc9902488b2",
      "role": "The three prompt slices every arm sent, byte for byte: the longest word-boundary prefixes of the frozen article corpus of at most 4,000, 32,000 and 128,000 tokens, each with its digest and token count, and the rule that cut them. Each request is a slice with a seeded ticket line before it and one instruction after it, as `client/prompts.py` builds it."
    }
  ]
}