← back to Local Model Leaderboard Watch

verification/e2e-proof.json

57 lines

{
  "intent": "Local defensive source fix: refuse config-only/incomplete/corrupt tensor files before model-load request. No runtime model operation requested or performed.",
  "risk_tier": "R1",
  "ticket": "TK-11401",
  "correlation_id": "TK-11401/tail/M-02912",
  "timestamp": "2026-09-11T00:44:30.033704+00:00",
  "base_commit": "4c16796",
  "environment": "Adopted local source repository; real review.sh exercised in temporary filesystem against mocked external commands; real model directory checked read-only by helper.",
  "preconditions": [
    "Real Qwen3.8 directory has config/index but zero shards",
    "TK11325 records intentional user-authorized removal",
    "No active related model ticket; prior owner notified M02912"
  ],
  "commands": [
    "python3 test-model-preflight.py",
    "bash -n review.sh",
    "git diff --check",
    "python3 verify-model-weights.py ~/.exo/models/mlx-community--Qwen3.8-27B-8bit"
  ],
  "assertions": [
    {
      "verdict": "PASS",
      "boundary": "adopted wrapper -> intercepted model API",
      "details": "41 offline positive/negative cases pass, including all22 supported dtype widths, forged shape size, malformed JSON objects, stub/missing/truncated shards, zero-sized and scalar tensors; no external requests."
    },
    {
      "verdict": "PASS",
      "boundary": "actual model filesystem",
      "details": "weight preflight REFUSED: missing or outside-root shard: model-00001-of-00006.safetensors; exit1"
    },
    {
      "verdict": "PASS",
      "boundary": "source adoption",
      "details": "Actual checked-in files exercised; bash -n and git diff --check pass."
    },
    {
      "verdict": "NOT_IN_SCOPE",
      "boundary": "download/model inference/scheduler",
      "details": "No download, wrapper invocation against real services, model swap, scheduler install/restart or inference validation required or performed. No claim of publisher authentication/engine compatibility."
    }
  ],
  "rollback": "git -C /Users/macstudio3/Projects/local-model-leaderboard-watch revert --no-edit 156aa46 6fe8096",
  "limitations": [
    "Validates local tensor layout, dtype/shape byte lengths and index references only; not publisher authentication, full safetensors conformance or inference-engine compatibility.",
    "Unknown dtype names fail closed and require explicit support review.",
    "Preflight cannot prevent concurrent deletion after check; runtime errors remain possible."
  ],
  "verdict": "PASS for local defensive source fix",
  "review_fix": "Acceptance reviewer proved F32 shape[1000] with4-byte offsets/payload incorrectly passed6afbb80. Followup rejects mismatch, validates malformed object schemas without traceback, supports zero-size tensors, replaces HOME override with scoped EXO_MODELS_DIR.",
  "reference": "https://github.com/safetensors/safetensors/blob/main/safetensors/src/tensor.rs (Dtype::bitsize verified September10,2026)",
  "adopted_commits": [
    "6fe8096",
    "156aa46"
  ],
  "source_adoption": "PASS: clean master based on4c16796, two reviewed local commits; actual adopted source harness41/41 PASS"
}