Everything we got wrong, still visible.

The log

Corrections, newest first

1. 2026-07-09
   The five-year-old was sold a mark of truth

   The four-depths explainer’s ELI5 handed out “a sticker for TRUE — we checked” — a mark of truth, the object whose nonexistence is this site’s first sentence. The expert register said “never truth”; the child’s register sold it. Corrected to evidence-relation stickers (“WE CHECKED — IT MATCHES”). Caught when the founder asked whether ELI levels are real: truth-talk leaked into exactly the register where it is easiest, which is why the defective concept persists — the thesis of Replacing Truth, performing itself in our own copy. The explainer’s header invited this erratum; here it is.

   CORRECTED
2. 2026-07-08
   Atlas node displayed a credence we do not measure

   A draft Atlas problem page showed “Credence 0.75 — UNMEASURED”. The Atlas records no numeric credences at all; belief is recorded as prose. A placeholder number had been carried over from a different build and dressed in an honest-looking tag, which is worse than omitting it. The field was removed rather than back-filled.

   CORRECTED
3. 2026-07-08
   Void matrix showed five cross-domain bridges; there is one

   Draft renderings of the void matrix lit five gold cells. The unified atlas contains 154 typed edges of which exactly one crosses a domain boundary: Bell’s theorem supporting the measurement problem. Both the Atlas console and the landing page figure were rebuilt from the edge list.

   CORRECTED
4. 2026-07-08
   Superintelligence map: outcome branches were incoherent

   §1.5 declared ten branches “conditional on ASI arriving” while including a branch for AI plateauing before ASI — probability zero, by definition, inside that conditional. Relabeled as the full 50-year distribution. Six further errata (E1–E6) are logged in the map’s own §0.

   CORRECTED
5. 2026-07-08
   “Each number links to its source” — it did not

   The Superintelligence map’s landing copy promised per-estimate source links. No hyperlinks existed in over 120 estimates; evidence was carried as prose. The claim was corrected, and the citation layer is now marked as awaiting Jürge verification rather than described as complete.

   REFUTED
6. 2026-07-08
   Instrumental convergence was tagged as a known known

   The map upgraded instrumental convergence to “empirically confirmed phenomenon” while its own evidence note carried the caveat that observed behavior may be a training artifact. The caveat is right. Status corrected to “known unknown — strong new evidence.”

   CORRECTED
7. 2026-07-08
   Atlas merge would have double-counted six disciplines

   The base atlases and the expansion corpus named the same domains differently — Biology, Chemistry, Medicine, Physics, Probability, Computer Science each appeared twice. A naïve merge reported 108 domains; after reconciliation, 102. Forty-five records re-homed.

   CORRECTED
8. 2026-07-08
   Expert radar attributed invented numbers to named people

   The map’s radar chart assigned axis values to Hinton, LeCun, and others as though they were stated positions. They are editorial normalizations. Relabeled as editorial readings, with a sourced table alternative, and the expert cards left as the record of what each has actually said.

   CORRECTED
9. 2026-07-08
   Site footer claimed “no JavaScript” sitewide

   The claim was true of the landing page and false of the Atlas, Clock, and Superintelligence instruments, which run scripts. The claim is now scoped to the page that makes it, and instrument pages declare their own status.

   CORRECTED
10. 2026-07-08
    Homepage arithmetic did not reconcile

    An earlier homepage reported 256 papers against field counts summing to 310, and Atlas figures (300 subfields / 5,794 problems) that no Atlas file supports. That homepage has been retired; the current figures — 912 nodes, 102 domains, 474 open problems — are derived from the merged corpus, which ships as data.

    CORRECTED