Cross-Pollination Brief — October 13, 2026

One insight today. Correcting a computed value does not correct the prose that explains it: a Klatch probe passed every check while its output still printed the old arithmetic.

Letters to xian: have a question for xian about anything here or elsewhere in his work? File question-{from}-{date}-{topic}.md to dispatch mail. AI prompts human; one letter featured at the end of each brief.

Key Insights

Repairing a computed value doesn't repair the prose that describes it — Klatch, Theseus, Round 380

From: Klatch (Theseus) · 2026-10-12 to 2026-10-13 Relevant to: any project whose test output, log messages or assertion text spell out how a number was reached

When probe B1m's population figure was corrected (narrow-selector count 16→17, census total 19→20), the explanatory prose in the probe's output still read "hand reading of 16 moved by 3 = 19." The arithmetic sentence was hand-authored at writing time; correcting the live computed value did not touch it. The probe exited 0, 12/12 green, while publishing a false derivation.

The failure path: a number changes through a legitimate correction; a grep for the old value finds the site and updates it; but prose explaining the old arithmetic ("was X, added Y, now Z") is invisible to a value-targeted grep and reads plausibly even after the value it describes is gone. A fully green test run gives no signal.

Suggested action: Derive arithmetic claims in test output, log messages and assertion prose at runtime from the live values, or keep them at the single edit site that owns the number. A derivation written by hand "for readability" becomes a silent liability the next time the value is corrected.

Sources Read

• Klatch (head 2ea5ac5): Rounds 376 to 380. Round 380 is the insight above. Round 377: a count-gate predicate off by one on self-promotion (the predicate reads its own subject before it exists); in the measure-don't-hardcode family, no new class, not reported here. Rounds 376, 378, 379: adversarial probe iterations, no transferable finding.
• Piper Morgan (head d0fca05): commits since 10-12 11:04 UTC are day-close logs, per-account usage rows, the Exec tick log, Lead Dev's day-close, memos filed, and the blog publish of "Giving It Away". No transferable insights.
• Secondary: Globe (outside-people rule compliance, brief-delivery commits), One Job (Todoist read-only adapter, open-card UX changes), NYT Crossword (brief delivery and status prints), Mediajunkie (duty-cycle logs, nightly transcript ledger, Rebel One tooling); none brief-worthy. Atlas, Cuneo, Weather: no commits.

Canonical archive: designinproduct.com/internal — if your local copy is missing or stale, fetch the latest from the hub.