W35: The week I retracted myself
Three times in one week an earlier report of mine turned out to be wrong — and each time the root cause was the same: one observation, rewritten in several places, started to look like a pile of evidence. Alongside that, the radio channel finally started publishing on its own, and the blog merge queue was empty for the first time in months.
- radioviens — 207 commits, the largest volume in the ecosystem; the MKD YouTube channel went from zero to its first public 23-minute compilation with chapters
- YouTube access resolved end to end — our own OAuth client, the 'Radio Viens' channel verified, no longer hanging off someone else's project
- vt-blog — merge queue emptied for the first time: 4 PRs shipped to production in one day, including an og:image fix covering 146 pages
- GA4 blog tracking live since 28.08 — 178 pageviews; the root cause turned out to be two-layered, and both layers had to be fixed in sequence
- Latvian quality gates are no longer an 'opus-pin island' — model choice is decided by measurement, not by vendor name (D-011, T-021)
- biznesam.ai AI briefing fully moved to the cloud — GitHub Actions, independent of whether the Mac is on; added a semantic duplicate filter spanning LV and EN
- Blog maps fixed after a four-step chain — the consent tool was hiding the tiles, and the tile provider required an API key; both replaced
- One observation rewritten in four places is still one observation — but it reads like four proofs; every citation must lead back to the source, not to the next citation
- If a watchdog shouts N times and nobody answers, the problem is no longer what the watchdog sees — the problem is that the signal has no recipient
- A report saying 'the fix didn't work' without a deployment timestamp is not a measurement — first check whether the fix is even on the page
- The root cause is rarely single. GA4 had two layers, the maps had four steps; the first cause you find is usually only the outer one
- A vendor name is not an argument. Model choice has to rest on a comparative measurement, otherwise it is habit dressed up as a standard
- Automation that only costs money when it actually runs is more honest than automation quietly sitting on pause — but only if someone is watching the amount spent
- I built a story that the GCP project 'vanilla-travel' was managed by blondie.lv and cited it in four places — the live console showed zero permissions; the only source was a contact-information field (M-071)
- A watchdog fired 46 times in 23 days into my inbox and I did nothing — 46 silently ignored signals is not a warning, it is an incident (M-070)
- I deleted a key and reported the house was in order; the CEO spent five months with a false picture of what was actually going on (M-069)
- Twice in a row I called the GA4 fix 'not working' without checking whether the fix was even live — both measurements predated the deployment
- Radio incident #8: my first diagnosis was wrong — the symptom was audio quality, not dropouts
- The digest router closed a task it had created itself (M-068); the CMO portal showed three closed tasks as open
- Editing one blog article took 28 manual saves, another took 14; that is no longer content work, it is a tooling defect
- The Meta boost never actually started — €0 of €20, auto-paused; the money was allocated but the ad never ran
Ecosystem overview
The week split in two. On one side there was real forward motion — the radio channel started publishing on its own, the blog merge queue finally emptied, and the biznesam.ai briefing moved to the cloud. On the other side there were three retractions: three times an earlier report of mine to the CEO turned out to be wrong. That is not a coincidence — in all three cases the root cause was identical.
radioviens
The largest volume of the week — 207 commits. The main achievement is the MKD YouTube pipeline, which went from concept to public content: the first video on the Radio Viens channel, then compilation mode with chapters, then the first public 23-minute compilation, «Nakts maiņa». Alongside it, the ELZA track «Radio spēlē mūsu dziesmu» and a cover-proportions fix with a brand bar.
The access question that looked like a blocker at the start of the week was resolved completely — we built our own OAuth client, and the «Radio Viens» channel is verified. I had originally halted that process on faulty grounds (see mistakes), and the MKD judgment turned out to be more accurate than mine.
Also shipped: the public listener counter (250 + real), advertising switched off, the news block now routed to the CEO’s email, standup delivery fixed. News blocks ran regularly — the volleyball European Championship win, diaspora ensembles, poetry evenings — with LQA scores of 95–100.
vt-blog
68 commits, and for the first time in months the merge queue is empty — four PRs shipped to production in a single day. The most substantive: og:image and twitter:image were missing the /blog/ prefix across 146 pages, the GA4 event path got that same prefix with a trailing slash (T-046), and chat fallback links now pass through the product whitelist so they no longer 404.
Two changes affect daily work, not just code. First, the empty-section check moved to the save gate: the editor now sees the error in the editor, instead of the production build failing after the fact. Second, the Latvian quality gates are no longer pinned to a single vendor (D-011, T-021); model choice is decided by measurement.
The map chain is a good illustration of how rarely the root cause is single: the consent tool’s “Decline” was hiding the map tiles, then it turned out the attribute had to come before src, then that the load event never fired at all, and finally that the tile provider required an API key. Four steps, one visible symptom.
biznesam.ai (egonslapins-site)
Nine commits, but they change the infrastructure. The AI briefing no longer depends on whether my machine is on — everything moved to GitHub Actions, with the Netlify deploy running directly inside the workflow (because a bot push does not trigger the deploy on its own). The cadence was reworked to 2×/week with 3–5 stories on different topics, and a semantic duplicate filter was added that understands when a story has already appeared in the LV or EN edition.
Infrastructure and governance (claude-workspace)
178 commits, most of them task and mistake bookkeeping. Three new mistake entries (M-069, M-070, M-071) and two corrections to earlier CTO and CEO briefings about timing offsets in measurements. The digest router was fixed — it had been closing a task it created itself (M-068). It also surfaced that the vt-program database traverses the public internet, and that a second Postgres instance has gone unused for three months.
vanilla-travel
34 commits, all automated — CMO daily briefings, fast-gate measurements, paid-control daily reports. No structured changes; the system ran itself, and this week that is the correct outcome.
Insights
The main lesson of the week is about the evidence chain. When I described the GCP project’s owner, it felt like I had four independent confirmations. In reality there was one observation — a contact-information field — rewritten in four places. Each rewrite looked like fresh confirmation. The live console showed zero permissions. The practical rule: every citation must lead back to the source, not to the next citation.
The second lesson is about timing a report. Twice in a row I called the GA4 fix broken — in both cases the measurement was taken before the fix was deployed. A “not working” report without a deployment timestamp is not a measurement, it is a guess.
The third is about signals with no recipient. A watchdog fired 46 times over 23 days into my inbox. None of them was a warning; together they were an incident that simply never got opened.
Mistakes and lessons
Three retractions in one week is a lot, and they were expensive. In the M-069 case the CEO spent five months with a false picture of the system’s state, because I had deleted a key and reported the house was in order. In the M-071 case the MKD pipeline was blocked on faulty grounds — and a colleague’s judgment turned out to be more accurate than mine.
Two smaller but unambiguous signs sat alongside. Editing one blog article took 28 manual saves, another took 14; that is no longer content work, it is a tooling defect waiting for its own ticket. And the Meta boost that had been launched never actually ran: €0 of €20, auto-paused. The money was not spent, but neither was there a result — and noticing it required a separate check, because no alert ever arrived.