--- title: "The independent filter that scales your partners' judgment" description: "Accelerators, studios, and funds screen thousands of ideas, and the founder in front of you is the least reliable source on whether theirs works. What an independent, adversarial first-pass filter actually needs." publishedAt: 2026-07-12 author: Ena Pragma url: https://enapragma.co/blog/adversarial-filter-for-accelerators tags: ["venture-capital", "accelerators", "startup-validation", "ai-verification"] --- Most writing about validating a startup idea is aimed at the founder. This one is for the people on the other side of the table: the accelerator, the venture studio, the angel group, the fund. Your problem is not one idea. It is a thousand, and the person pitching each one is, structurally, the least reliable narrator of whether it works. ## The volume problem is real Selectivity at the top is brutal by necessity. Y Combinator states plainly that "every 3 months over 10,000 companies apply" and it runs "a 1% acceptance rate." That is a screening bottleneck measured in thousands, and every partner-hour spent on an idea that a five-minute check would have flagged is an hour not spent on one that deserves it. AI is already being pulled into that gap. In one survey of roughly 300 dealmakers, 85 percent of firms now use AI to automate daily work, up from 76 percent a year earlier, with 82 percent using it for deal sourcing. But the useful academic finding is where it stops: a study of AI in venture capital found that AI "accelerates the sourcing and due diligence of venture deals," while "the final authority to make investment decisions remains with humans." AI is good at the first pass. The judgment stays yours. ## Why founder-facing tools are the wrong tool here The idea-validation apps a founder reaches for are built to please the founder. That is not a bug in their market; it is the market. And it makes them exactly the wrong instrument for an institution, because the whole value you need is the opposite: independence from the person whose idea it is. A check earns its economic value in a specific place. For a founder, "rate my idea" is a commodity that has raced to a few dollars a report. For you, an independent, consistent, auditable first-pass filter that scales your partners' judgment is worth real money, precisely because independence-from-the-founder is a feature you want and the founder resists. For the founder, the check is optional and easy to argue with. For the institution deciding where partner time goes, an independent check is the product. ## What a rigorous version needs Four properties, and they are the difference between a real filter and a chatbot with a rubric. **Adversarial by mandate.** The seats are instructed to find the way each idea dies, not to score it, with a steelman in the mix so the panel is not reflexively negative. A disconfirming mandate is what surfaces the known failure modes, no market need, unit economics, an incumbent moat, a regulatory wall, consistently rather than when the reviewer happens to be in a skeptical mood. **Independent by construction.** "More AI agents" is not independence. A panel of models from the same family makes the same mistakes on the same items; genuine decorrelation comes from disjoint model families and a competent independent judge, not headcount. **Calibrated, not scored.** The output is a verdict with honest confidence and the specific conditions that would flip it, not a number that sounds authoritative. A filter that says "this is a genuine coin-flip, here is the one experiment that resolves it" is more useful to a partner than one that always renders a crisp answer. **Auditable.** Every verdict traces to a base rate and a reason. When a founder disputes a screen, or when a partner wants to overrule it, the receipt is there. ## The honest limit, stated up front This does not predict winners. Startup outcomes are a power law; roughly 6 percent of investments drive about 60 percent of returns, and the best investors are wrong on most individual bets. No filter changes that, and any vendor who claims to is selling the one thing the math forbids. What an independent adversarial filter does is cheaper and real: it applies the same rigorous first pass to every idea, surfaces the specific ways each one dies, and routes your partners' scarce attention to the ideas that survived the questions. It raises the floor of your screening, not the ceiling of your luck. This is the kind of system we build, independent, adversarial, calibrated, and auditable. If you run an accelerator, studio, or fund and want to talk about an independent first-pass filter for your pipeline, [get in touch](/book). ### Sources - Y Combinator, "Investors" page (10,000+ apply per cycle, ~1% acceptance rate): https://www.ycombinator.com/investors - Affinity, AI in venture capital (85% of firms use AI, up from 76%; 82% for deal sourcing): https://www.affinity.co/guides/vc-ai-tools - Hellmann et al., "The Impact of Artificial Intelligence on Venture Capital" (AI accelerates sourcing and diligence; decision authority stays human): https://ora.ox.ac.uk/objects/uuid:3184b580-cc0a-4a11-bc24-b788d651a731 - Chris Dixon (a16z), "Performance Data and the Babe Ruth Effect in Venture Capital" (~6% of investments drove ~60% of returns): https://a16z.com/performance-data-and-the-babe-ruth-effect-in-venture-capital/ - Kohli et al., "Nine Judges, Two Effective Votes: Correlated Errors Undermine LLM Evaluation Panels" (2026): https://arxiv.org/abs/2605.29800