Michael Darius Eastwood · Independent AI alignment researcher
Published
A register is worth what its sources are worth. If half the DOIs, arXiv numbers, patent identifiers and institutional URLs cited alongside the nineteen rows fail to resolve two years from now, the register has quietly become a story that used to be evidence. The monitor is the cheap piece of infrastructure that keeps this from happening without anyone noticing.
Once a week, a script iterates every row in the register and re-resolves every primary source it names. Each check produces one of four outcomes. Green: the identifier resolves and returns content matching the archived title and authors. Amber: it resolves but the metadata has changed, which usually means a versioned arXiv upload or a moved DOI target. Red: it fails to resolve, either because the URL has broken or because the identifier no longer maps to the expected object. Grey: the target is behind an intermittent gateway and the check is inconclusive. A summary is committed to the repository each week; the run history is public.
A red row moves to a quarantine section on the register page, with its status stamped and its last-known-good date recorded. It is subtracted from the headline count in the same commit. The programme's public phrasing therefore falls, sometimes by one row, sometimes by more, in step with the reality of what a reader can still open. If the source recovers under a corrected identifier, the row returns; if not, it stays in quarantine and eventually is removed with a note that it failed source-checking. A register that can go down is an instrument. A register that never moves is a press release.
Two amber events on arXiv identifiers whose version numbers had incremented, requiring the row to be updated to the current version. One red event on a press URL that had been consolidated into a paywalled aggregator; the arXiv preprint remained resolvable, so the row's primary source was rewritten to the arXiv identifier and the press URL demoted to supplementary. No permanent red events yet, which is roughly what a young register with strong primary sources should expect and is not a claim the monitor is doing nothing. It has already changed rows, in public, on the record.
It cannot verify content beyond title and authors; a paper could resolve to a target whose contents have been silently amended, and the monitor would report green. It cannot detect a source retraction that has not been reflected in the identifier's metadata. These are hostile cases the human reviewers are responsible for. What the monitor buys is that no row survives quietly on a dead link, which is by far the commonest failure mode of long-lived reference lists and the one that a Sunday-afternoon script can eliminate. The next hardest failure mode, silent tampering, gets human attention because the monitor has cleared the ambient noise.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.