Michael Darius Eastwood · Independent AI alignment researcher
Published
Every claim in this research programme is published alongside the conditions that would prove it wrong, on a public dashboard, with live status. One of those conditions has already fired: an early scaling figure failed cross-architecture replication and was retracted. This article explains why the dashboard exists and what it costs.
Most research communication optimises for the strongest defensible presentation of favourable results. The dashboard inverts this: it hands any hostile reviewer a target list. Run the blinding protocol yourself. Check the nineteen register sources. Find the pre-2024 document. Build the counterexample system. The bet is that in a field drowning in unfalsifiable claims, the scarce resource is not confidence but checkability, and a programme that survives published kill-conditions accumulates a kind of credibility that no amount of assertion can buy.
The alpha 2.24 figure was attractive, quotable and wrong; replication produced 0.49, sub-linear. Retracting it cost a striking number and bought something better: the demonstration that the kill-conditions are real. Every grant application that carried the old figure was corrected. The retraction is now the first thing on the dashboard, deliberately, because a reader who sees a fired kill-condition treats the unfired ones as meaningful.
Publishing kill-conditions is cheap to start and expensive to honour, which is why it is uncommon. It binds your future self in public. But the entire architecture of this programme is a wager that verification beats persuasion, and a falsification dashboard is simply that wager applied to our own claims rather than only to everyone else's.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.