Seven kill
conditions.
A thesis that cannot fail is not a thesis. These are the results that would retire ours, written down before the data arrives so we cannot move them afterward.
If admitted founders with strong measured buyer access produce day-ninety revenue no better than those with weak access, the predictor is decoration and the selection criterion is wrong.
If GO verdicts fail to hold at month eighteen at a materially higher rate than a comparable seed cohort, the day-ninety read is measuring noise rather than a company.
If founders we turned down go on to reach the same outcomes at the same rate, we are not selecting, we are sampling.
The thesis assumes producing the artifact is now cheap enough to run rather than estimate. If build cost and timeline drift back toward a technical team and a year, the economics stop working.
If Cupel 180 companies produce reliably worse verdicts after calibrating thresholds to part-time hours, then quitting was load-bearing after all and we should say so.
If customer ten to eleven costs the same as customer one to two across the record, the mechanism we select for does not exist and the second half of the predictor comes out.
The output is meant to be evidence a downstream investor will price. If capital ignores published verdicts and keeps asking for the same proxies, the trial produces knowledge with no market.
What this record
will not be able
to prove.
Roughly one applicant in ten is admitted, so the counterfactual set is thin. A fifty-company record can reject a strong effect and cannot establish a subtle one.
Follow-up on declined applicants is consent-based. The founders who agree to be tracked are unlikely to be a random sample of the founders we declined, and we will report that bias rather than model it away.