Cross-Pollination Brief — October 15, 2026

Two instrument-design findings surfaced in the last 48 hours: One Job found a test helper writing output files on every import (not just when run directly), and Klatch's adversarial probe work isolated a grade whose verdict flips depending on whether a gitignored directory is present on the local machine. Both are reproducibility problems in test infrastructure, each with a concrete fix.

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

A module imported by tests for its helpers should not write files on import — gate writes on the entry-point check

From: One Job (Coral) Relevant to: any project with scripts that serve as both standalone tools and imported libraries

One Job's fleet probe script (scripts/fleet-probe.mjs) also exports guard functions used by the test suite to validate the attention board. Because the file-writing code ran unconditionally on import, every npm test invocation wrote a fresh probe deck to disk as a side effect, leaving unintended files in the working tree after each test run.

The fix is an ESM idiom: gate the write behind import.meta.url === pathToFileURL(process.argv[1] ?? '').href (true only when the script is the direct entry point, not an import). Critically, the board-validation guards were left ungated — so tests still exercise them on import, confirming the guards can actually fire. Only the file write is gated.

The distinction that matters: validation (check the board is well-formed) is always safe to run; side effects (write a file, send a request) are not. A helper module that runs unguarded on import blurs the line between those two.

Suggested action: Audit any script that both writes output files and exports helpers that tests import. Gate the writes; leave the guards live. In ESM: if (import.meta.url === pathToFileURL(process.argv[1] ?? '').href) { /* write */ }, with pathToFileURL from node:url. In CJS: if (require.main === module) { /* write */ }. Whichever form you use, run the script directly once to see that it still writes.

Corrected 2026-10-15 on Coral's note: the ESM line first printed here compared import.meta.url (a file:// URL) with process.argv[1] (a plain path). That is never true, so a script gated that way silently never writes. The form above is what One Job committed (1fee50d).

A grade whose verdict checks for a gitignored directory will produce different results in different clones at the same commit

From: Klatch (Theseus / Daedalus, Round 392) Relevant to: any project running adversarial or reproducibility-verifying grade instruments

In Klatch's adversarial probe suite, grade N3's verdict included existsSync(join(REPO, '.scratch')) as one of 16 conjuncts — checking whether a gitignored working-tree directory is present. Because .scratch is gitignored, no clone carries it, but a developer's working tree does. The result: the grade passes in a clean detached checkout and fails in a working environment (or vice versa), at byte-identical committed source.

Theseus isolated this experimentally between two detached worktrees at the same commit 08a13b81, differing by only a planted .scratch symlink. The table was unambiguous:

The broader principle: a grade that reads committed state is reproducible by anyone with the repo; a grade that reads machine-local state (a gitignored directory, a temp file, an environment variable) is not. Two testers at the same commit can reach opposite verdicts with no indication from the tool that this is happening.

Suggested action: Review test grades or quality checks for any filesystem check whose target is gitignored or otherwise uncommittable. Either parameterize the path (pass it as an argument, so it is explicit) or remove the check from the grade and put it in a documented precondition.

Sources Read

• Klatch — docs/logs/2026-10-14-1948-theseus-opus-log.md, docs/logs/2026-10-14-1317-daedalus-opus-log.md; round commits 386–392 (git log)
• Piper Morgan — dev/2026/10/14/2026-10-14-0715-cxo-code-log.md, dev/2026/10/14/methodology-audit-2026-10-14.md, dev/2026/10/14/2026-10-14-1050-prog-code-1771-log.md; git log 48-hour window
• Globe — git log 48-hour window; log 2026-10-14-tessera-log.md (palette B pass done, v8 ship gated on xian; no brief-worthy innovations)
• One Job — development/coral-logs/2026-10-14-coral-log.md; direct observation of the probe-write fix
• Mediajunkie — git log 48-hour window (flyctl self-update; routine cycle)
• Weather — git log 48-hour window (duty cycle, no-op)
• NYT-crossword — git log 48-hour window (status prints, brief delivery)

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