{
  "intro": [
    "My choices are **#1, #3 and #6**. These are scalable commercial hypotheses, not demonstrated billion-dollar valuations. No single measurement establishes market size, manufacturability, customer adoption and defensibility together."
  ],
  "rows": [
    {
      "row": "#1 — PFAS-free genus",
      "proof": "One independently qualified, purchasable lead meeting a buyer’s complete two-phase coolant specification. Lock the molecule/formulation before testing; use the electrical, phase, materials and boiling gates already organized in `docs/WETLAB_SOP.md`.",
      "controls": "Incumbent reference and failed-spec controls on the same instruments; identify static versus electronic permittivity explicitly.",
      "expected": "A candidate-specific qualification report showing simultaneous compliance. The imported target is ε `<2.5`, GWP `<250`, and boiling window approximately `45–65°C`, from `imported/3prok/00_TOP3_IDEAS_RANKED.md`; the buyer must confirm its actual specification.",
      "kills": "No purchasable lead passes the combined specification, or qualification relies on substituting electronic polarizability for static permittivity/breakdown. One failure kills that lead, not every possible member of the genus."
    },
    {
      "row": "#3 — Pentane+DMC",
      "proof": "A controlled bare-copper boiling comparison of the exact mass composition against pure fuel and the commercial reference, with phase and electrical prerequisites passed first. Use the repo’s bare-fluid protocol, without enhancement hardware.",
      "controls": "Same geometry, pressure, surface preparation and fluid conditioning; randomized replicate runs. Pure fuel separates a blending benefit from ordinary boiling performance.",
      "expected": "A reproducible boiling curve and CHF with uncertainty, showing a useful advantage while retaining acceptable electrical and phase behavior. The repo’s modeled deployment threshold is `50 W/cm²` in `src/screening/enhancement_stack.py`; it is a proposed gate, not a demonstrated result.",
      "kills": "Phase separation, unacceptable electrical behavior, or no meaningful thermal benefit over controls. A regenerated `3.50×` FoM does not rescue a failed physical comparison."
    },
    {
      "row": "#6 — discovery engine",
      "proof": "A blinded external surface-tension benchmark with grouped holdout by chemistry and data source, evaluated through the predictor without dictionary-answer leakage. This is the strongest compute-first proof.",
      "controls": "Shuffled-target model, simple baseline, duplicate/source-overlap audit and explicit out-of-domain reporting before unblinding.",
      "expected": "Held-out R² exceeding the already advertised `0.75` threshold, with uncertainty and error by chemistry class. Source of that claimed threshold: `BEST_IDEAS.md` row 6.",
      "kills": "R² fails the advertised threshold, leakage explains the score, or the shuffled/baseline controls perform similarly. The engine might remain useful, but the advertised predictive moat fails."
    }
  ],
  "coda": [
    "The engine can already perform much of the third proof’s computation. What is missing is a properly independent, locked benchmark—not another run against its own lookup anchors.",
    "I did not select #8 despite its attractive recurring-service potential: its existing drift model fails conservation. That must be repaired before its forecasts support a commercial argument."
  ],
  "target_date": null,
  "source": "gpt-6-astra read-only audit of this repository, 2026-09-07 (C1, item 3)"
}