zae.life ~ $ log correcting-a-claim-added-two

  Correcting one unsupported claim added two false ones
  2026-08-13  ·  attestation

  claimed  ──
  A claim found to outrun its evidence gets corrected, and the
  document is then more accurate than before. Review removes bad
  claims.

  observed ──
  In one entry, across three revisions: two claims outran their
  evidence in the draft — one attributing a motive nobody observed,
  one asserting something never measured. Correcting the second
  introduced two more, and both were false rather than merely
  unsupported: that the account in question was the longest of three
  and the most specific, when a different account was each. A fourth
  followed in the message reporting those corrections, calling two
  instances a pattern. Each was written while explaining why claims
  must not outrun evidence.

  attested ──
  That in this document, on this occasion, a correction introduced
  claims that were not there before, and that they were worse than
  what they replaced. Not that revision generally does this — one
  document is not a rate. Not that four is the count: it is the number
  found, by two readers who opened the file, and no search was
  exhaustive. And not that this entry is outside its own finding. It
  is another revision of the same material, written by the same
  author, and has not been audited by anyone but him.

The correction was the ordinary kind. A reviewer pointed at a sentence claiming something about the intelligence and care of three accounts, which nothing in the check had measured. That was true, and the sentence went.

What replaced it said the account in question had been the longest of the three and the most specific about the source. Neither is so. A different answer was the longest; a different one again gave the most specific figures. Removing one claim that reached past its evidence produced two that were simply wrong, and they read as more precise than what they replaced — which is presumably why they were written.

There is a tempting explanation: that saying why something is so requires generalising, and generalising is how a claim leaves its evidence behind. It fits all four instances. It is also untested, produced by the same faculty that produced the four, and appears here only because leaving it out would be its own kind of tidying.

What can be said without it is narrower and more useful. A revision is not a repair applied to a fixed object; it is new text, and new text carries new claims. The verification that cleared the previous version does not transfer to it. This is uncontroversial for code — no one merges a second diff on the strength of the first review — and is almost never done for the prose that explains the code.

Counting the fifth generation, this one: it has been read once, by the person who wrote the other four.