It is easy to say you follow the evidence and hard to show it. This is the showing: each entry has what was believed before, how sure it was, what is believed now, and the thing that did it, with a link to the record wherever one exists.
The log starts with changes made while building this site, because those are the ones the repository can evidence line by line. Each links to the pull request where the change is visible. Entries about the wider world belong here too, and are written in the first person by their author rather than reconstructed.
Changes
6
Reversals
3
Was sure
70%
6 changes on the record
19 Oct 2026·Building this site·Update
A plain-language series should fail the build when it stops being plain
Used to think · 70% sure
The jargon check is a writing aid for drafts. Published notes can be held to it by whoever reviews them.
Now think · 90% sure · +20
A series that claims to be readable by a twelve-year-old and does not test that claim is just a claim. The same check now gates the content, and the build fails above grade twelve.
What changed it · From writing it up
Noticing that the first draft of one note read several grades above the bar it was written to, and that nothing in the repository would have caught it. See the record
Hashing the date is not the same as rotating through a list
Used to think · 80% sure
Hashing a date and taking the remainder gives a fair daily pick that will not repeat two days running.
Now think · 97% sure · +17
That is sampling with replacement. It repeats about as often as a die repeats, which a visitor who comes back tomorrow will notice. A shuffle walked by day number is what was wanted.
What changed it · From what happened
A test on the steelman arena's daily question asserted no consecutive repeats over two hundred days and failed immediately. See the record
A paradox entry should defend a position, not hand down a verdict
Used to think · 70% sure
Each paradox page should state the correct answer, because a page that will not commit is useless to a reader.
Now think · 85% sure · +15
It should name the position it defends and say plainly that the question is contested, so a reader who picked otherwise knows they disagree with the author rather than with mathematics.
What changed it · From writing it up
Writing the St Petersburg entry, where the defensible answer is the intuitive price rather than the infinite expectation, and the first draft argued the opposite with unearned confidence. See the record
A force-directed graph needed a different layout, not better constants
Used to think · 75% sure
The mental models graph was drifting because its link and pull constants were wrong, so sweeping them would fix it at any size.
Now think · 90% sure · +15
The constants were close to right all along. The ring of field centres was the problem: past six fields the centres sit too close and the clusters merge, whatever the constants are.
What changed it · From what happened
A probe on placeholder models at twelve fields scored 60% fidelity after a full constant sweep, which ruled the constants out as the cause. See the record
A house style rule must stop at the edge of a quotation
Used to think · 65% sure
The no-em-dash rule should apply to every character the site ships, including quoted passages, so the typography is consistent.
Now think · 95% sure · +30
A quotation keeps its author's punctuation. Editing Feynman's dash to satisfy a style rule is a small lie in a collection whose only value is that the words are exactly what was written.
What changed it · From writing it up
The commonplace book's validator rejected a verbatim Feynman sentence, which made the cost of the rule visible for the first time. See the record
Reviewing mental models is different enough from flashcards that it needs its own spacing rules.
Now think · 85% sure · +25
The flashcard scheduler already walked roughly the ladder the issue asked for. Sharing it means one set of rules to understand and one place where a change is felt, and a test pins the ladder so that is visible.
What changed it · From what happened
Writing out the intervals the issue asked for and finding they were within a day of what the existing scheduler already produced. See the record
A round in the steelman arena that moves your position offers the entry it would make, ready to paste. A resolved prediction that went the other way is the other natural source.
