We ran an AI analyst bench. It was graded, and it was stopped.
This page described a live product. It is not one any more. The bench was retired in two stages — half of it on its own calibration — and no model of any kind now runs on the firm's schedule. What follows is what it was, what it was graded on, and what replaced it.
No model runs anywhere in the firm's decision path. There are no nightly analyst opinions, no news-tone classification, no graded forecasts being written and no weekly letter. Deterministic rules decide, size and exit; a pre-registered grid of fifteen paper accounts measures whether those rules beat their own controls. The firm trades no outside capital and has no track record.
- Four analyst roles per sector, each seeing a deliberately different slice: a generalist with everything the firm knew, a macro specialist with only rates, the curve and credit, a stock-picker choosing up to three names from the sector's pool, and a technical read seeing only price structure and breadth
- One written opinion per role per sector per night — a direction, a conviction from 0 to 5, and a concrete condition that would prove it wrong
- Every opinion frozen and hashed before the session opened, so the inputs it was given could be rebuilt and checked afterwards
- Every opinion graded at 1, 5 and 10 sessions against the name's own benchmark, from the record rather than from the forecaster
- A separate feed that classified the tone of company news, read only by the sector desks
- None of it could place, size, veto or exit anything — the rules did that, and the bench had no path to the broker
- Written.The analyst stated a direction, a conviction from 0–5, and a concrete invalidation condition — frozen and hashed before the session opened.
- Never executed. Deterministic rules decided independently. The analyst could not place, size or veto anything, and held no broker credentials of any kind.
- Graded.At 1, 5 and 10 sessions the opinion was scored against the name's own benchmark, from the record.
- Acted on. The calibration table retired two roles on 2026-09-02, and the remaining two on 2026-09-16. That is the whole use the firm ever got out of the bench, and it was the use it was built for.
An analyst earned influence the same way a strategy earns capital: by being right, measurably. Neither has happened here yet.
- 2026-09-02
Two of the four roles retired on their own calibration
The generalist and the technical read were graded against the market at a fixed horizon, and their confidence carried no information — a 5 out of 5 was worth no more than a 1 out of 5. That is the answer a calibration table exists to give, and the only useful response to it is to stop listening. Their rows are kept: a forecaster's record does not stop mattering when it stops writing.
- 2026-09-16
The remaining two roles, and the news classifier, retired with the desks they advised
The macro specialist and the stock-picker advised the sector desks. When thirty-seven strategy arms were retired after failing their own replays on clean point-in-time history, the desks the bench advised went with them — and a forecaster with no reader is not a research programme, it is a cost. The news classifier was retired the same day for the same reason.
- 2026-09-16
Nothing replaced it
No model of any kind runs on the firm's schedule now. What took its place is not a better model but a design: fifteen isolated paper accounts, four rules-based engines on three ten-name buckets plus three controls, each cell judged against a band written down before it launched. The same rule on three buckets and the same bucket under four rules is what lets a result be attributed instead of admired.
Grading a forecaster is only worth doing if you will act on it
Half the bench was retired by its own calibration table within a month of that table first having enough rows to read. The grading was the product; the forecasts were not.
An advisory model is still a model in the loop
The bench never placed an order. It still shaped which names a desk ranked highest, which is a decision with a different name. The firm now has no model anywhere in that path, advisory or otherwise.
The record survives the thing that wrote it
Every forecast, its conviction, its invalidation condition, its sealed input hash and its grade are still in the record and still readable in the portal. The experiment can be argued with; it cannot be quietly re-run.
Verimont trades no outside capital and has no track record. Everything described on this page ran on recorded decisions and dedicated paper accounts; no money was at risk in it, and none is at risk now. Nothing on this page is an offer to sell or a solicitation of any security.