Michael Darius Eastwood · Independent AI alignment researcher
Published
Someone paying close attention to the record notices that earlier versions of documents contained different values. The convergence count that once said thirty-one now says nineteen. A scaling exponent of 2.24 that once appeared in drafts is now flagged retracted. A garbled valuation figure was quietly withdrawn. The observation is correct and the conclusion the observer sometimes draws from it is wrong. Numbers changing under audit is not the failure mode; numbers never changing under audit is.
Three things, in three separate corrections, all documented. The register count fell after a six-lane audit applied the rules of one external event per row, one primary source per row, and moved unsourced candidates to a published pending list. Alpha 2.24 was withdrawn when replication across architectures failed to reproduce super-linear scaling, producing a robust estimate near 0.49 in the opposite direction. A misread valuation, quoted in draft as a figure of about ninety-six billion dollars framed as if it were an IPO price, was found to be a garbled combination of a confidential S-1 filing and a $965 billion valuation figure, and was withdrawn from the corpus with the accurate context noted where relevant. Each correction is on the record with its date and its trigger.
Every load-bearing value in this programme derives from a canonical facts file. When one changes, the change happens in one place. A truth gate then lints the corpus for stale appearances of the old value before the new commit is allowed to land. Documents pick up the new value on the next regeneration; nothing is hand-updated across dozens of files, and nothing is quietly forgotten. The gate has forbidden-pattern entries for exactly the numbers that used to appear and no longer should. If a retracted figure returns bare in any file, the gate blocks the commit.
One of two things. Either the audits have never caught anything, in which case the audits are not working, or the audits have caught things and the record has hidden them, in which case the record is not credible. A programme that spends eighteen months on this material with zero visible corrections has done something wrong and is asking the reader to trust it about which of two possibilities applies. The current record spends the credibility that visible correction buys, on purpose. The falsification dashboard opens with the retraction; the audit trail on the count reduction is public; the withdrawn valuation carries a note where it used to appear.
"You changed your numbers, so nothing you say is trustworthy." The move relies on the reader treating any change as evidence of failure regardless of direction. It is worth naming: this framing rewards corpora that never correct anything and punishes those that do. The relevant question is not whether the numbers changed but whether the changes went in the honest direction, were documented on the record, and moved by the mechanism that would move them again if a new problem surfaced tomorrow. The answer to the objection is not defence; it is precisely the mechanism the objection was built to attack. That mechanism is the thing the reader is being asked to trust, and the corrections are its receipts.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.