AINEWS
The open source project asktheboard treats AI advice like a clinical trial: each prediction gets a timestamp before the outcome is knowable, then a calibration score when reality resolves. The first graded bet lost, and that failure is itself the proof the mechanism works.
The first graded bet this system published came out wrong. That is the most useful fact about it.
An open-source project called asktheboard treats AI advice the way a clinical trial treats a drug. Before the answer is knowable, the system records a dated, falsifiable prediction, with the per-seat probabilities of each model on the panel. When reality arrives, it reconciles a Brier score (a standard measure of how well calibrated a probability forecast was, where lower is better) for each seat and a calibration curve for the panel. The point of the system is not to be right. The point is to be auditable.
The first resolved example is anchored in examples/2026-06-jobs-report.md. The board pre-registered a multi-seat call on the June 2026 U.S. jobs print. Each seat registered a probability vector alongside its reasoning, and a dissent vector, showing disagreement across seats, was logged with the minute. When the print came in, the consensus call was wrong, and the system published the per-seat Brier score on the resolution date.
A second minute, examples/open-minute.md, is still in flight. That is the ledger part of the design: bets are public, dated, and they age. The project landing page and PyPI package frame the system as an open pre-registration ledger, not a chat product.
What actually gets recorded on each decision: the user's stated prior (the assumption the human is bringing in); each seat's stance and probability, with disagreement made explicit; a dated, falsifiable prediction; and on resolution, per-seat Brier scores plus a panel-level calibration curve. The mechanism is borrowed from pre-registration in clinical trials and the social sciences, where the integrity of an inference lives in the timestamp, not the personality of the predictor.
The interesting move here is not the personas or the chat surface. It is the timestamp. Most 'AI advisory boards' ship with persona-flavored debate and a chat window and little else. Few require the call to be locked in, with a timestamp, before the outcome is knowable, and even fewer publish a reconciliation score when reality resolves. asktheboard does both, on a public, git-anchored minute.
The caveats matter. One graded bet is not a track record. The June 2026 U.S. jobs print specifics referenced in the worked example, including the headline nonfarm-payrolls number and the unemployment rate, should be checked against the Bureau of Labor Statistics release before any specific number is treated as canonical; the project itself is a solo-developer open-source launch with no third-party validation of the Brier reconciliation. The demo's outcome is hand-supplied, and git-anchored timestamps are a fragile attestation layer: signing, repo availability, and host durability all have to hold for the integrity claim to scale.
The structural claim still survives those caveats. The market for AI advisors is filling with persona-flavored debate tools. The live question is whether advice that was not anchored before the outcome was knowable can be trusted. Pre-registration plus a dated, falsifiable prediction is the integrity layer most of those tools skip. The launch has been discussed in a Hacker News thread.
The portable test, then, is not whether the AI is right. It is whether the AI will be auditable. Ask the next AI advisor you meet two questions. Did it publish the prediction, dated, before the outcome was knowable? When reality resolved, did it publish the score? If the answer to either is no, you do not have advice. You have a demo. If yes, you can read the receipt. The receipt is the product.