Absence of Evidence Is Evidence Against

The standard view — and why it doesn't apply here

The phrase "absence of evidence is not evidence of absence" comes from epidemiology and philosophy of science. It describes a real and important asymmetry: the failure to detect something doesn't mean it isn't there. A doctor who hasn't run a test can't conclude from the missing result that the patient is healthy. The absence is uninformative precisely because the test wasn't performed.

This is the correct view in the general case. But it doesn't apply to structured documents — and conflating the two is one of the most common sources of overscoring in any evaluation system.

When you submit a pitch deck, grant proposal, or business plan, you are not conducting a passive experiment where certain results may or may not have been measured. You are making an argument. You control what goes in. Every omission is a choice — or a gap.

What structured documents imply

A pitch deck that doesn't mention customer validation hasn't "not yet been asked about it." It has been asked — implicitly, by the genre itself. Every investor who reads a pitch is asking: has this founder talked to real customers? The deck is the founder's opportunity to answer that question. A deck that provides no answer has answered: not yet.

The same logic applies to every document that makes a structured case. A grant application omitting a risk mitigation section hasn't been denied the space to address risks — the genre provides that space, and the applicant didn't fill it. A product spec with no definition of launch criteria hasn't been prevented from specifying them — it has chosen not to.

In each case, the absence is informative. The reader knows what should be there. The absence tells them it isn't.

How Bayescore treats absence

Bayescore's adversarial pass is specifically designed to surface this kind of absence. The supportive pass extracts what is present: named customers, documented demand signals, quantified market size. The adversarial pass looks at what is missing: a document that describes a product without naming a single customer contact, a plan that claims distribution without specifying a channel, an application that asserts novelty without citing any prior art.

When the adversarial pass finds that a high-weight predicate has no supporting evidence in the document, it returns a confidence score near zero. Not because the founder lacks customers — they may have talked to dozens. But because the evidence isn't in the document, and the document is what's being evaluated.

This is a precise and deliberate design choice. The evaluation is of the document, not of the founder. The document either contains the evidence or it doesn't. If it doesn't, the score reflects that. The fix is not to argue about the score — it is to put the evidence in the document.

The practical consequence

A pitch deck that scores 34/100 with customer validation failing is not a verdict on the business. It is a verdict on the document. The question "have you done customer discovery?" is answered by what you wrote, not by what you know. If you've talked to forty customers but your deck doesn't mention a single one, the score is 0.0 on that predicate.

This surfaces a genuine problem that informal evaluation consistently misses. Founders who have done the work but haven't documented it receive the same investor skepticism as founders who haven't done it at all — because the investor, like Bayescore, can only evaluate what's in front of them.

The right response to a failing predicate is almost never "but I did do that." It is: "where in the document does it say so?" If the answer is nowhere, that's the edit to make — before the pitch, before the application, before submission.

The Bayescore self-evaluation

Bayescore evaluated itself on IS(startup, launch_ready) — a domain that checks customer validation, demand signals, go-to-market, and domain expertise. The score was 46/100, Grade D. Five predicates failed because the document submitted didn't contain evidence for those things. The score was correct. The full predicate-by-predicate breakdown is public at bayescore.com/self-eval.

A score that flatters you when the evidence is absent is not useful. A score that accurately reflects what the document contains — including what it's missing — is. Absence of evidence is evidence against. That's not a bug in the scoring system. It's the system working.