We spent a week trying to prove our own product doesn't work.

Since March we have published a multi-model AI consensus: five models, a directional call, an agreement score on the front of every page. In late July we scored 942 company-level calls against what the prices actually did, measured against the base rates you would get by simply guessing.

The direction was decoration. Our calls were right 39.2 percent of the time. Guessing the single most common outcome every time was right 39.9 percent. Broken out, bullish calls hit 32.0 percent against a 31.5 percent base rate; bearish hit 26.8 percent against 28.6 percent. Every bucket landed at chance, and the agreement score added nothing on top of it.

So we removed it. The directional signals are gone and the daily journals are paused.

One thing survived. When all five models agreed, the stock moved 2.93 percent over the following week. When they disagreed, 3.98 percent — roughly a quarter less movement under agreement. It held when we compared each stock only against its own history, and it strengthened as agreement rose.

Xavier cannot tell you which way something will move. It appears to know when something will be quiet. That result rests on 16 trading days across 82 companies: promising, not proven. We are testing it forward before it becomes anything you can buy.

The written analysis was never the weak part. On 27 March we argued that the Strait of Hormuz was an insurance story rather than an oil story, that war-risk premiums would reprice by 0.5 to 1.5 percent of hull value, and that the increase would stick regardless of how the diplomacy went. Premiums moved from 0.25 percent to multiples of that, softened on the June memorandum, and rose again.

But we kept re-covering that story and we got worse at it. By late April we were publishing a war-risk baseline of 0.25 percent as though it were current. It had been stale for seven weeks. That is not bad reasoning. It is no memory: each brief was written as if the story had just started.

Of 4,296 forward-looking items across our briefs, roughly one in six could be checked at all. The rest told you what to watch rather than what would happen.

Every claim we publish from here carries four things: a named public indicator, its value today, the threshold that would confirm the call, and a date by which it resolves. Published before the outcome is known. Scored afterward in the open — the ones we get wrong shown beside the ones we get right.