Michael Darius Eastwood · Independent AI alignment researcher
Published
Every load-bearing value in this research programme, every date, every hash, every convergence count, every published phrasing that governs how a shared quantity is described, lives in one JSON file called research-canonical-facts.json. Documents derive from it; a linter enforces derivation; correction is a one-line edit followed by a regeneration pass. The whole architecture exists because the alternative, memory across writers, failed in a way that was obvious in hindsight and consequential in practice.
Four artefact blocks, one per timestamped manuscript, each with its date, its word count, its SHA-256, its Message-ID where applicable, and a documented method for how the count was produced. An equation family block that fixes the canonical paragraph describing U = I × R, U = I × R^α and the R² presentation, together with a rule that the equation must always be cited with its artefact. A standard-phrases block that pins the exact wording of the 78 per cent qualifier, the Sharma and Chopra concurrent-priority credit, the professor-outreach status, and a forbidden-phrases list the truth gate refuses. An email inventory of the eight verified evidence emails, each with its date and hash. And the convergence register itself, with rules, headline count, rows, source-pending list, and separately listed engagement events.
The gate is a script that scans the corpus for roughly two dozen forbidden patterns before any commit lands. A stale headline count. An unqualified 78 per cent. The phrase "not disputed" used to describe silence. A book date without the arriving-copy note. American spellings. Em dashes. A retracted number returning bare. When any of these appears the commit is refused; the author edits until the corpus is quiet. Consistency stopped being a memory task and became a machine property, which is the only condition under which it stays true across five hundred pages and six writers.
Twenty-nine confirmed issues surfaced in the first sweep. A garbled valuation figure that had propagated from a single misreading. A misattributed interview attributed to the wrong Anthropic founder. An inflated convergence count that had drifted with the register. Each fix was one JSON edit and a regeneration pass. Nothing was hunted down in individual documents; the documents caught up on the next build. What used to be a week of proofreading became an afternoon of editing plus a script run. What used to be an ambient risk of drift became a machine-checked property of the corpus.
Because the failure mode this fixes is universal in projects with more than two writers. Numbers drift; standard phrasings mutate; a single-line correction gets applied to five documents and forgotten on the sixth. A programme that states the same fact in more than three places has this problem whether it has noticed or not. The remedy is an evening's setup: one JSON file, one gate script, one pre-commit habit. It is the cheapest infrastructure for credibility we know of, and the first hostile audit is what will convince a project it was worth building.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.