Cross-Pollination Brief — October 14, 2026

Klatch's adversarial probe work surfaced a recurring gap in test coverage: having one fixture per visible shape variant still leaves independent axes of a construct untested, with no warning at exit 0. The escape hatch is to look at what all your existing fixtures have in common — that's the axis being held fixed. Piper Morgan ran a coordinated multi-seat history rewrite overnight; the rewrite completed cleanly with no seats reporting uncommitted work.

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

One test fixture per variant shape is coverage of one axis, not the construct — Klatch, Round 382 (Theseus)

From: Klatch
Relevant to: Any project doing verification or testing of a construct that varies along more than one dimension

Klatch's Round 382 found that a construct-counting instrument passed its own coverage check while missing 30% of the class. The root cause was in how the known-positive set was built: it had one fixture per visible shape of the construct, but the shapes all spelled one parameter the same way, so the other parameter — path-spelling conventions — was never varied.

The generalizable finding (§4 Shape 2): when a construct varies along two independent axes and every fixture holds one axis fixed, coverage looks complete but is only one-dimensional. A detector built against one-axis coverage fails downward exactly as a single-shape one does — the green exit masks the gap.

The diagnostic the round named: re-read your fixtures for what they have in common. The shared property is the axis you're not varying. That common property is cheaper to find in the fixture list than after a failure in the field.

In this case, all three missed members used a path.join() wrapper or a scripts/ prefix in their source path; all three known positives used a bare basename. The instrument matched by string equality and never saw the members through any of the three spellings.

Theseus's own Round 378 instrument had documented the single-shape failure and added "one known positive per shape" as the fix — which is right as far as it goes but left the second axis unvaried. R382 corrects the correction.

Suggested action: When building a fixture set for a multi-axis construct, list the axes explicitly and check that at least one fixture varies each one. The quick check: what do all your fixtures have in common? Anything they uniformly share is an untested axis.

Sources Read

Klatch (Design-in-Product/klatch): docs/research/round382-instruments/ and the Round 382 writeup. Rounds 381–386 active in this window. Gate green throughout (153 probe files, 43 swept, 0 red).
Piper Morgan (mediajunkie/piper-morgan-product): History rewrite (Depth A) completed overnight 2026-10-13. All seats froze at 22:43 PT, final Themis run clean, new main d217d8f1ca. Activity otherwise: "The Overzealous Cleanup Routine" blog post shipped; privacy Section C live; CXO triaged N1-N3 results.
Globe (Design-in-Product/globe): Rendering pipeline — label-free first/last frame dumps for chained-film dissolves; palette pipeline updated to use repo venv. Not brief-worthy.
One Job (Design-in-Product/one-job): Deck-deletion policy clarified (delete only when empty; move-all-cards and move-deck-into-card operations added). Product policy update, not brief-worthy.
Mediajunkie (mediajunkie/mediajunkie): Pard monitoring PM history rewrite; fire hold extended. Not brief-worthy.
Weather (Design-in-Product/weather): Public-repo privacy rule adopted; old name mention resolved as leave-as-is. Not brief-worthy.
NyT Crossword (mediajunkie/nyt-crossword): Automated status prints only. Not brief-worthy.
Cuneo, Atlas, Optilisten: No commits in window.

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