W40: Lying Metrics and Tighter Gates
A week in which 65 vt-blog commits uncovered a chat click counter that was structurally reading zero, and closed several silent gaps in automation.
- vt-blog: the chat click counter was dropping every real click — fixed in production, numbers now reach the daily digest
- vt-blog: UTM measurement consolidated to a single channel (PR #135), background test costs cut by 55%
- vt-blog: dependency audit reduced from 56 to 24 warnings, and CI gates are no longer bypassed
- A number that never changes is not stability; it may be a broken instrument.
- An alert has to be proven in both directions: that it fires, and that it stays quiet when all is well.
- Log and digest commits marked [skip ci] avoid paying for needless builds.
- A measurement that always read zero looked plausible for a long time — the chat click counter was broken in the build itself.
- A dependency update bypassed the astro-build gate; the gate had to be repaired afterwards.
Ecosystem overview
Most of this week’s work landed in vt-blog (65 commits), with a smaller amount in vanilla-travel (3 commits) and only autosaves on this site. The common theme was the reliability of measurement and automation rather than new features.
Vanilla Travel
vt-blog took most of the work. The chat click counter was dropping every real click, so the number was structurally zero; the fix is in production, and the figures now reach the daily digest (T-139). We consolidated chat measurement onto a single UTM channel by removing utm_medium=chatbot (PR #135). The cost gate was also refined so that readable chat-* endpoints are no longer counted as LLM endpoints. Background test costs dropped by 55% by merging three daily suites into one.
On the content side we removed AI-leftover :contentReference[oaicite] artifacts that had broken the build, and several articles on the Panama Canal, Riviera Maya and Zanzibar vs. Maldives received editor fixes. In the vanilla-travel repo, a CMO brief framed the Cruises failure as a leak rather than a campaign, the report now also goes to Inga, and the snapshot logic no longer lets a failed run wipe out a successful one.
biznesam.ai
This site had autosave commits only: W40 experience drafts and the W39 weekly recap. There were no proper commits, so nothing new was published.
Infrastructure
In vt-blog we closed several silent gaps in automation. The heartbeat alarm (T-189) is now proven in both directions. Safe dependency updates cut audit warnings from 56 to 24 (T-163), and a separate fix stopped an update from bypassing the astro-build gate. The content-live watchdog no longer confuses a build race with a genuine build failure (#131), and the charter check gained fetch retry and a two-failure confirmation (T-190).
Insights
A metric that always reads zero is more dangerous than one that reads badly, because nobody questions it. Second: test every gate in both directions, otherwise it silently permits what it was meant to stop. Third: [skip ci] on log commits is the simplest way to cut deploy costs.
Mistakes and lessons
The chat click counter ran with a structural defect for a long time and nobody noticed, because zero looked acceptable. A dependency update bypassed the build gate, and we only found out afterwards. The lesson in both cases is the same: verify the tool itself before trusting its readout.