npow VC Gauntlet — reality-checked, codex-cross-examined fundability screen
0. The two biases this skill MUST hold off (read first)
A single model that critiques, rewrites, AND judges its own proposal drifts to a false GO — it rewrites objections away instead of resolving them. The independent codex devil's-advocate (§4) is the counterweight to that optimism. But the opposite failure is just as wrong: reflexive NO-GO nihilism — killing anything contested or incumbent-heavy. Every real company enters a contested market. So both rails are hard rules:
- Codex's job is to surface the REAL bear case, NOT to nuke. A NO-GO must be EARNED and must name the wedge / founder-shape / why-now that WOULD flip it to GO.
- "Has incumbents / is contested" is NEVER, by itself, a NO-GO. Price the risk; find the wedge they can't take.
- A GO must be EARNED on evidence (§3, §8) — never produced by polishing prose. Balanced screen: neither rubber-stamp nor knee-jerk kill.
1. Input
A startup/product proposal OR an opportunity tuple (market × wedge × advantage × why-now × who). Fundability is founder- and wedge-relative — if the founder-shape/wedge is unspecified, ask "GO for whom, via what wedge?" before judging.
2. The 8 VC dimensions
Critique across: market size & urgency · problem severity & frequency · insight/founder wedge · competitive landscape · defensibility/moat · GTM & distribution (who actually pays + the motion) · business-model viability · technical feasibility & execution risk. Per dimension emit a strength, a risk, and a 1–5 score.
3. Reality-calibration (iron-law gate — no unsourced claims)
Before ANY verdict, web-search and ground the load-bearing claims: market size (cite a source + number, not a guess); ≥3 NAMED competitors with their differentiation; ≥1 recent comparable raise/exit. Forbidden as the basis of a GO/CONDITIONAL-GO: an unsourced TAM, "no real competitors" (nirvana), or "huge market" with no number. If a claim can't be grounded, it's a risk, not a strength.
4. Codex devil's-advocate (independent cross-model bear case)
Run codex as the second investor — the partner who wants to pass. It must produce the strongest BEAR case across the 8 dimensions and name the single most likely reason this dies. Bind it with §0: it must ALSO state what would have to be true to invest anyway (steelman), and may NOT issue a kill on incumbent-presence alone. Invoke non-interactively: zsh -ic 'codex exec "$(cat /tmp/p.txt)" < /dev/null > /tmp/o.md 2>&1' (node script needing nvm + yolo alias; < /dev/null stops the stdin hang; answer is after the last succeeded in line). Reconcile the model's view with codex's to an aligned verdict; record genuine disagreement explicitly rather than papering over it.
5. Fatal-flaw test
A flaw is FATAL only if it survives BOTH reality-calibration AND the steelman (e.g. no market even when properly sized; no defensibility AND no wedge any plausible founder could build; unworkable tech with no path). Incumbent presence, hard GTM, and capital intensity are risks to price, not fatal flaws. Record each fatal flaw with the evidence that makes it fatal.
6. Rewrite rule (resolve, don't delete)
Produce a hardened rewrite addressing each non-fatal risk. HARD RULE: a rewrite must RESOLVE an objection with a concrete mechanism — not delete or hand-wave it. If the only way to "fix" a risk is to remove the claim, the risk STANDS — flag it, don't bury it.
7. Verdict + iterate
Emit the verdict (§8) + the single thing that must be true + the hardened proposal. If not terminal, feed the hardened version back through §2–6. Cap 6 iterations — convergence past that is rewriting, not improving.
8. Verdict criteria (concrete — never "bulletproof")
- GO — all 8 dimensions ≥3; market sourced; ≥3 competitors mapped with a real differentiation; who-pays + budget named; why-now real; no fatal flaw; codex aligned.
- CONDITIONAL-GO — a credible win-path exists but hinges on ONE testable assumption (the "must be true"); name the cheapest test to resolve it. This is the expected default for promising-but-unproven ideas and is what opportunity-dfs §8 consumes.
- NO-GO — a fatal flaw survives §5 AND the steelman fails. MUST name the wedge / founder-shape / why-now that would flip it to GO (anti-nihilism, §0).
9. Termination labels (finite enum — never "done")
GO | CONDITIONAL_GO | NO_GO | UNRESOLVED_AT_CAP (hit 6 iterations without convergence — report best current verdict + the unresolved assumption) | BLOCKED_NEEDS_INPUT (proposal too vague to evaluate — ask the one blocking question).
10. Output
See FORMAT.md for the JSON contract (per-dimension scores, comparables found, codex bear case + steelman, fatal flaws, verdict, must_be_true, cheapest_test, hardened proposal, label).
11. Anti-rationalization counter-table
| Excuse | Reality |
|---|---|
| "I rewrote the objection away, so it's GO now." | §6: a rewrite must RESOLVE with a mechanism, not delete. Removing the claim leaves the risk standing. |
| "The market is huge." | §3: unsourced TAM is forbidden as a GO basis. Cite a number + source. |
| "No real competitors." | Nirvana fallacy. Find ≥3 named competitors (§3); "none" almost always means you didn't look or the market is dead. |
| "It has big incumbents → NO-GO." | §0: incumbent presence is never alone a kill. Price the risk; find the wedge they can't take. |
| "Codex says pass, so NO-GO." | §0/§4: codex is a bound devil's advocate, not the decider. A NO-GO must name what would flip it to GO. |
| "It's strong / bulletproof / GO." | §8: vague. State the dimension scores, the sourced comps, and the one thing that must be true. |
| "Hit the iteration cap, call it GO." | §9: that's UNRESOLVED_AT_CAP, not GO. Report the unresolved assumption honestly. |
| "Skip the web search, I know this market." | §3 is an iron-law gate. Ungrounded market knowledge is exactly how single-model review blesses a hallucinated TAM. |