Michael Darius Eastwood · Independent AI alignment researcher
Published
When six parallel work streams write about the same research, numbers drift. A convergence count becomes 15 in one document, 17 in the next, then 28, 29, 31 and 32 as authors copy from different snapshots. A publication date mutates. A percentage loses its qualifier. Each individual error is small; together they are fatal to a programme whose slogan is don't believe, verify, because the first inconsistency a reader finds discredits every number they have not checked.
The fix is borrowed from software engineering. A single canonical facts file holds every load-bearing value: dates, counts, hashes, gaps, standard phrasings. A gate script then lints the entire corpus against roughly two dozen forbidden patterns before any commit: stale counts, unqualified percentages, imprecise timestamp language, banned causal claims, superseded dates, and, after one embarrassing audit, em dashes and American spellings, which violate the programme's house style. A violation blocks the commit. Consistency stopped being a memory task and became a machine property.
In its first sweep the gate and the audit behind it surfaced twenty-nine confirmed issues, including a garbled valuation figure, a misattributed interview, a false-priority implication against other researchers, and a register inflated by duplicate rows. All were fixed corpus-wide in a day, because fixing meant editing one file and re-running a script.
Any project that states the same facts in more than three places has this failure mode. The remedy costs an afternoon: one JSON file, one grep-based gate, one pre-commit habit. It is the cheapest credibility infrastructure we know of.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.