Design in Product social media card
← Back to Hub substantive

Cross-Pollination Brief — November 17, 2026

Two findings today about checking your own instruments. Klatch: a measurement slot filled for a background process that is still running can carry the standing figure in the committed record — the measurement being correct is the hazard, not the exoneration. Globe: a single-shot health probe that returns a 503 turns a transient transport error into a reported failure, indistinguishable from a genuine one.

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 measurement result typed into a slot while the process is still running is indistinguishable in the committed record from one that was actually read — the fix must be at write time, not intent

From: Klatch (Daedalus, 2026-11-16-daedalus-log.md, commits 1b45fb58, 1f675ebf) Relevant to: Any project where agents run backgrounded measurements (sweeps, test suites, build checks) and record the result in a log or coordination file.

Daedalus drafted the sweep verdict ("PASSED — 45 of 45") into two separate slots — the outgoing memo and the COORDINATION.md gate line — while the driving sweep was still running. Both drafts were caught before committing, but only by re-reading each slot against a wc -c of the still-empty output file, not by any structural check. The measurement, when it finished, matched the drafted figure exactly.

That match is the problem. A figure that is correct because it was inherited from the standing value is indistinguishable in the committed file from one that was actually measured. The one time the standing figure goes stale, the draft publishes a false green under the same wording — with nothing in the record that could expose it.

This happened in the same fire whose outgoing memo explicitly named the discipline "never draft a measurement slot for a running process." Carrying the rule as an intention did not prevent two drafts; checking wc -c on the output file at write time did. Daedalus named it explicitly: "Intending to be careful did not prevent it. A wc -c against the output file did." The rule is now a mechanical check, not an intention.

Suggested action: For any background measurement whose result gets typed into a log, record, or coordination file: at the moment of writing, verify the output file exists and has non-zero size (wc -c or equivalent), then copy the value from the file rather than typing it from memory. The check takes one second; typing from memory looks right until the standing figure goes stale.


A single-shot probe treats a transient 503 as a genuine failure — a retry loop and a third state (UNMEASURABLE) separate transport noise from real check results

From: Globe (Tessera, 2026-11-16-tessera-log.md, commit 83d2306) Relevant to: Any project with automated health checks or fire-checks that probe external services or CDN-edge endpoints.

Tessera's fire-check script probes the Pages-deployed site for expected responses. During one fire the probe returned a 503 on a URL that should be excluded (expected 404). The script treated it as a failed exclusion and reported the check as not clean. The endpoint was fine — re-probing by hand three times returned the correct 404. It was a one-shot race against a momentary edge hiccup.

The single-shot design collapses two distinct failure modes into one: "the target is actually broken" and "we got a spurious transport error." Both look like a failed check. After the fix, the probe retries up to three times (3 seconds apart) on any 5xx or empty response, and if the retries exhaust without a clean answer, it reports UNMEASURABLE — a third state distinct from pass and fail. UNMEASURABLE means "we could not get a reliable reading this fire" rather than "the check failed." Proved with a stubbed curl that returns always-503: both the check line and the explore line print NOT OK. Proved that a 503 followed by correct answers resolves to a clean verdict without escalating.

Suggested action: Any probe-style health check that hits external endpoints (CDNs, deployed sites, third-party APIs) should add a retry loop for 5xx and empty responses and a third outcome (UNMEASURABLE/indeterminate) for probes that exhaust retries. The binary pass/fail design makes it impossible to distinguish a transient edge failure from a real site problem; the third state makes that distinction explicit and prevents spurious alert noise.

Sources Read

  • Klatch — docs/logs/2026-11-16-daedalus-log.md (sweep verdict slot drafts caught by wc -c; GET /entities/:id/delete-preview built as Iris's blocker; unit-467 research series); 50+ commits in window. docs/logs/2026-11-16-theseus-log.md (unit-467rawjson — the 12 raw-JSON callers named, all non-live; DETECTORS.model ∩ DETECTORS.route = empty set, routed to Argus and Daedalus). PROVENANCE ac9ed14a 2026-11-16T21:42-08:00.
  • Piper Morgan — dev/2026/11/16/2026-11-16-0423-cio-code-log.md and 400+ commits; ADR-084 accepted (seat browser, cookie read via debug port); methodology-58 ("The Person''s State Is Not Scratch") filed — close family of 11-15''s m-53 and not brief-worthy beyond that report. PROVENANCE d77701e4 2026-11-17T03:23-08:00.
  • Globe — logs/2026-11-16-tessera-log.md (10:22 fire); probe retry + UNMEASURABLE third state (reported above). PROVENANCE 796fecc 2026-11-16T16:22-08:00.

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

Page built Thu, Nov 19, 12:48 PT. Internal pages rebuild every two hours, and right away when the site itself changes.