Cross-Pollination Brief — October 18, 2026

Klatch's adversarial testing rounds continued through R405–R410, producing one sharp finding: a test instrument committed into the corpus it searches will contaminate itself in both directions, but the red direction gets caught and the green direction gets believed.

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

When a test instrument is committed into the corpus it searches, the green contamination is more dangerous than the red

From: Klatch (Theseus — Round 410, commit 0803db46)

Relevant to: Any project that commits search-instrument code into the same codebase the instrument scans.

When a corpus-search test tool is committed to the repo it searches, its own source content becomes part of the corpus. The CONTROL_BODY fixture — a "known negative" whose job is to not appear anywhere in the corpus — found itself after the instrument file was committed. That is a red result and it is obvious: the instrument immediately reports an unexpected match, the developer investigates, and the fix is structural (exclude the instrument's own directory from the search using a self-directory seam derived from import.meta.url).

The more dangerous direction is green. In the same round, a prediction about a known positive was confirmed by a member the round itself had just committed as evidence — rawprobe…mjs:3 landed in the corpus that same session. The instrument reported GATE true, prediction LANDED. That passed without investigation, because a confirmed prediction stops being a question.

The asymmetry: a false negative (red) gets examined. A false positive from self-contamination (green) passes. Both are caused by the same mechanism — the instrument's own committed artifacts appearing in its own scan — but only one of them raises a flag.

Three fixes applied together in R410: (1) derive the instrument's own directory from import.meta.url and pass it to the corpus reader as an exclusion; (2) grade the self-exclusion seam both ways in a gate leg (seam off → control appears in corpus and only there; seam on → control vanishes; known positive survives both, because a seam that hides a real member is a widening, not an exclusion); (3) key predictions on the inherited corpus, not the current working tree, so evidence committed by the round cannot confirm a prediction made before the round started.

Suggested action: If a repo has corpus-search instruments — graders, scanners, analyzers that run against the project's own source — check whether those instruments exclude their own file paths from the search. A one-line exclusion (import.meta.url → exclude the instrument's own directory) is sufficient; the gate leg that verifies the exclusion works in both directions is the part that requires thought.

Sources Read

• Klatch — docs/logs/2026-10-17-1947-theseus-opus-log.md (Round 410 session log); commit series R405–R410 in the 48-hour window.
• Piper Morgan — commit dd7c2850c4 (m-51 adoption note, docs/internal/development/methodology-core/methodology-51-A-BOUNDED-SEARCH-IS-NOT-A-TOTAL.md): Lead adopted a git grep pre-push check for pin text before editing pinned test strings; operational adoption, not new methodology.
• Globe — handoff second cut (10-16/17), render pipeline complete on prequel v3; no methodology to report.
• One Job — applied 10-17 brief (census blind-spot check across findLastIndex, match/exec, split(x)[n], and Swift optionals; no fail-open found); Pilot 44 Beta App Review approved. No new methodology.
• Mediajunkie — backup-age guard: added a conditional third retry only when tmutil status reports Running = 1, to distinguish "actively backing up" from "broken" on Time Machine health checks. Operational fix; platform-specific.

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