Goal — To generate valid trade signals: indications you can lean on because they were examined before they were believed. The benefit lies in that order — a claim about the market gets settled rather than argued, including, and especially, when it is settled against one’s own idea.

Challenge — Examining an idea costs a multiple of what having it costs, and almost always ends in a no. Then comes the more uncomfortable half: a system that computes continuously has to be able to show afterwards what it computed at a given moment. Without that, every later evaluation is a story — and one the teller adjusts without noticing.

Current state — Built and in operation is a bench that settles a claim about the market before any money is on it: shuffled data, a sample held back until the end, a positive control that finds planted effects and never invents one. 16 hypotheses fixed in writing beforehand, from 5 mechanism families, have been through it — 16 decisions that would otherwise have stayed opinions, and the bench let none of them through. That closes the search in freely available data and moves the work on: 3 forward-running experiments are open, their read date fixed in advance — fixing the date beforehand means you cannot choose your result afterwards. Also built is the path outward — computed cards go to the operator as a message — and a chain of signed snapshots evidencing what was computed and when. The operator is the only recipient; there is nothing to subscribe to, no access and no passing on. live

What part AI plays — In the judgement none: this is statistics — permutations, correction for multiple testing, a held-back sample. No model decides whether a hypothesis passes. In the building, everything: the bench itself, the ingestion, the delivery path and the signature chain are developed by AI agents. Their remit stops explicitly short of strategy, signal and snapshot — the boundary runs between building and judging, and it is written down.