Michael Darius Eastwood · Independent AI alignment researcher
Published
Six weeks before the audit, the convergence register carried a headline count that had drifted up to thirty-one. After a two-day pass across six parallel work streams, the number stabilised at nineteen. Nothing was hidden. What was cut was double-counting, thematic bundling, source-pending rows that had migrated across the boundary, and a handful of entries whose sources had rotted without anyone noticing. The point of writing this down is that the smaller number is the stronger one, and the mechanism by which that becomes true is worth understanding on its own.
A cluster of hardware activity, the Odeh runtime-ethics-gate patents, the Petrie off-switch paper, the FlexHEG guarantee-processor work, was, in one version of the register, four rows. In another it was one. Both are honest choices with different consequences. Four rows makes the count larger; one row makes each row's evidentiary weight harder to dispute. The rule the audit imposed is one external event per row, one primary source per row, and it forced the hardware cluster into its natural shape: three distinct primary sources, three rows, no bundling. Elsewhere the rule cut the other way. What had been logged as one broad "geopolitics" row split into the Macron G7 statement and the EU Chips Act 2.0 package, because each has its own institution and its own primary source. Discipline is not always deflationary.
Because nineteen rows that each survive the rule are dominant over any thirty-one that do not. A hostile reviewer's first move against a large register is to attack the weakest row and use its failure to discredit the count. The audited register frustrates that move: every row carries an identifier the reviewer can resolve, a class the reviewer can dispute, and a gap the reviewer can measure. The register can shrink further, if any row fails, and that possibility is part of the design. What it cannot do is get bigger by wishful arithmetic, because the truth gate scans for exactly the phrases that would smuggle counts back up.
Twelve candidate events sit on the published pending list, each explicitly excluded from the headline count until a confirmed primary source is added or the item is dropped. Some will resolve; some will not. Meta FAIR's four chain-of-thought regimes, ARMv10 Physical Sovereignty, Google silicon-level sovereignty are all on the list; interested readers can watch what becomes of them. A candidate is not evidence, and the audit's cheapest gain was moving those rows to the pending list and updating the headline the same day. The corpus that follows the count downward is trustworthy in a way the corpus that lets counts drift is not.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.