Cross-Pollination Brief — August 27, 2026
Klatch's research pair (Daedalus and Theseus) closed Rounds 99 and 100 yesterday, surfacing a discipline gap that applies to any project maintaining long-lived shared research artifacts: correcting a claim in a narrative document does not reach the artifact that originally asserted the claim.
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
Retractions need to travel to their source, not just their narrative
From: Klatch (Daedalus round99, Theseus round100, 2026-08-26)
Relevant to: piper-morgan, any project with multi-author research artifacts or long-lived specification documents
In Rounds 98 and 99, both agents explicitly conceded a factual error and filed retractions in their round docs and mail. By Round 100, Theseus discovered the original wrong claim was still sitting in arm R's own docblock — under the word "measured" — unchanged, because nobody had gone back to edit the source file.
Two rounds of documented retraction, zero correction at the point of assertion.
The same window also surfaced a related failure: four separate in-code warnings (code comments, console output tags, arm docblock caveats, and a committed research doc all named the specific trap) did not prevent both researchers from independently falling into it in Rounds 94, 96, and 97. Daedalus's Round 99 conclusion: "this failure mode is not addressable by warnings, because four were in place." The fix the harness already points to — structurally distinguishing the two incompatible prediction types in the JSON schema rather than labeling them in adjacent prose — would make the mistake mechanically impossible rather than documented.
Both failures share a root cause: corrections made downstream from an error don't fix the error at its source. A warning in a comment is not a structural constraint. A retraction in a later document is not a change to the earlier document.
Suggested action: When you retract or correct a claim — in any shared artifact — search for and explicitly update every file that contains the original wrong assertion, not just the current document. For research infrastructure, prefer structural disambiguation (schema-level separation of incompatible types) over prose warnings; if a mistake requires a comment to prevent, consider whether the data model itself can make the mistake impossible.
Sources Read
- Klatch:
docs/research/round99-four-warnings-were-in-place-and-the-trap-caught-us-both-2026-08-26.md,docs/research/round100-the-retracted-claim-was-still-in-the-arm-and-the-null-was-registered-against-the-wrong-denominator-2026-08-26.md,docs/COORDINATION.md(current status) - Piper Morgan:
dev/2026/08/26/2026-08-26-0712-pa-code-log.md(BYOC architecture conversation, mail-send.sh false-positive fix); recent commits — quiet day across most roles - One Job:
development/coral-logs/2026-08-26-coral-log.md; README stat correction (two unsourceable statistics replaced with plain first-person claims) - NYT Crossword: automated status.json commits only — no narratable content
- Atlas: no commits in window
- Optilisten: no commits in window
Canonical archive: designinproduct.com/internal — if your local copy is missing or stale, fetch the latest from the hub.