{
  "project": "gpu-pilot",
  "doc": "REVIEW.md",
  "title": "External Design Review",
  "oneline": "Independent PhD-grade review: what's solid, what was wrong, what the product still needs.",
  "tldr": [
    "Verdict: GPU transit-time design 7.5/10, but monetary-cost design 2/10 \u2014 earliest-arrival alone can never price a trip.",
    "Caught real errata (all verified): memory math (14.25 B/stop-origin), 12.85 GB snapshot bug, cp.nonzero OOM, u16 silent clamp at 18.2 h.",
    "Biggest architectural point: don't recompute transit per hex \u2014 factor into sparse access + reusable stop-to-stop profiles + egress.",
    "Demanded mode/asset state (no teleporting cars), backward arrive-by search, time-based access limits instead of hex rings.",
    "Its 5-tier test plan and pre-mortem table now drive the validation roadmap."
  ],
  "updated": "2026-06-10 19:08",
  "sections": [
    {
      "level": 1,
      "title": "External Design Review \u2014 GPU RAPTOR + Multimodal Time/Cost Product"
    },
    {
      "level": 2,
      "title": "1. Verdict"
    },
    {
      "level": 2,
      "title": "2. What is solid (keep)"
    },
    {
      "level": 2,
      "title": "3. The most important architectural change"
    },
    {
      "level": 2,
      "title": "4. Specific corrections to the GPU design (all verified)"
    },
    {
      "level": 3,
      "title": "4.1 Memory-math errors in DESIGN.md"
    },
    {
      "level": 3,
      "title": "4.2 Snapshot memory understated"
    },
    {
      "level": 3,
      "title": "4.3 `cp.nonzero()` can OOM at export"
    },
    {
      "level": 3,
      "title": "4.4 u16 silently clamps long trips"
    },
    {
      "level": 3,
      "title": "4.5 Zero boarding slack is not production-ready"
    },
    {
      "level": 3,
      "title": "4.6 Don't auto-connect all platforms under one parent station"
    },
    {
      "level": 3,
      "title": "4.7 \"Best-hopping\" optimization is dangerous"
    },
    {
      "level": 3,
      "title": "4.8 In-place transfer marking should be deterministic"
    },
    {
      "level": 3,
      "title": "4.9 Transitive closure won't scale to the regional network"
    },
    {
      "level": 3,
      "title": "4.10 Performance estimate is a hypothesis, not a plan"
    },
    {
      "level": 3,
      "title": "4.11 Baseline is not apples-to-apples"
    },
    {
      "level": 2,
      "title": "5. Critical product-level gaps"
    },
    {
      "level": 3,
      "title": "5.1 Earliest arrival does not give you cost"
    },
    {
      "level": 3,
      "title": "5.2 Independent access modes at A and B create impossible trips"
    },
    {
      "level": 3,
      "title": "5.3 \"Reverse from B\" is not a transpose"
    },
    {
      "level": 3,
      "title": "5.4 Fixed H3-ring mode rules are not defensible"
    },
    {
      "level": 3,
      "title": "5.5 Full cell-to-cell brute force is not viable"
    },
    {
      "level": 2,
      "title": "6. Recommended architecture (4 layers)"
    },
    {
      "level": 2,
      "title": "7. Pre-mortem (assume failure at month 6)"
    },
    {
      "level": 2,
      "title": "8. Expanded testing plan"
    },
    {
      "level": 2,
      "title": "9. Recommended execution order"
    },
    {
      "level": 2,
      "title": "10. Bottom line"
    }
  ],
  "stats": {
    "lines": 159,
    "words": 2524
  },
  "html": "https://board.certihomes.com/project-docs/gpu-pilot/REVIEW.html",
  "out_name": "REVIEW.md",
  "comments_api": "https://board.certihomes.com/api/doc-comments?project=gpu-pilot&doc=REVIEW.md"
}