{
  "schema": "techfuel-evidence-release/1",
  "case_id": "lmstudio-ollama-startup",
  "revision": "1.0.0",
  "status": "corrected",
  "decision_date": "2026-11-05",
  "observation_date": "2026-08-26",
  "canonical": "/data/llm-server-compare/#august-26-load-timers",
  "scope": "Four normalized dense-model rows, on one RTX 5080. Unloaded-to-completed-short-reply, resident decode and resident server TTFT describe different events. Client-visible TTFT was not measured.",
  "change": "The original CSV's first request begins before model loading in Ollama and after a separate CLI load in LM Studio. This first versioned timing export states those boundaries and derives comparable short-reply estimates without changing the measurements.",
  "affected_fields": [
    "request_includes_load",
    "unloaded_to_short_reply_approx_s",
    "resident_decode_256_median_approx_s",
    "resident_server_ttft_median_ms"
  ],
  "supersedes": null,
  "prior_reference": "/data/llm-server-compare-2026-08-26-lmstudio.csv",
  "checksum": "sha256 of UTF-8 file bytes normalized to LF",
  "files": {
    "json": {
      "stable": "/data/lmstudio-ollama-startup-2026-08-26/timing.json",
      "pinned": "/data/lmstudio-ollama-startup-2026-08-26/revisions/1.0.0/timing.json",
      "sha256_lf": "48813feeec90a9375abf381b53a567ca8db6bb5f38989369e883233f89c08de0"
    },
    "csv": {
      "stable": "/data/lmstudio-ollama-startup-2026-08-26/timing.csv",
      "pinned": "/data/lmstudio-ollama-startup-2026-08-26/revisions/1.0.0/timing.csv",
      "sha256_lf": "80bae738a287d91ec5aba4f90266d6abace4ab8d8fc95e93ac7b69edd41dcf7a"
    }
  }
}
