W30: Autonomy needs brakes
The week the radio station learned how to stop itself, the first client got a plan, and we were reminded that a silent failure costs more than a loud one.
- Radio Viens got a news gate that closes on failure — no fresh, verified content means no broadcast, rather than a broadcast with errors
- Silver Fox: the first client now has two pilots, a four-phase roadmap and a technical setup; the work surface is built on code of ours that already runs
- Wrote a standing rule into the ecosystem — for anything a customer reads, the model is chosen on quality first and price only after
- Moved the Vanilla Travel chat classifier to a cheaper model with a separate provider pin for personal data — cost down, privacy kept
- You judge an autonomous system not by how well it runs, but by how it stops
- A silent failure costs more than a loud one — the week's most expensive problem was not a crash, it was three songs that simply never played
- An alert needs delivery confirmation and a repeat damper, or it turns into noise and the real warning goes unheard
- Alerting cried red twice in one week where nothing was actually broken — the fastest way to teach yourself to ignore alerts
- Three already-approved songs silently never made it to air because a signature format didn't match what the gate expected — no error, just absence in rotation
- In the programme generator, the quality check deleted every day except the first — fixed, but the class of bug is worth remembering
Ecosystem overview
This was a week about how systems stop. Not about new features — about gates, breakers and signals. The radio station got a contract defining when it is allowed to broadcast and when it is not. Monitoring cried wolf twice and was rebuilt. Model selection got a rule that puts price second. And alongside all of it, the first real client got a first plan.
Radio Viens
The heaviest work of the week — nearly a hundred commits — went into the station’s autonomy. It started from a simple question: what does a system that broadcasts on its own do when the content isn’t ready? Until now the answer was “broadcast anyway”. Now the news gate closes on failure: not fresh, not verified — not on air. Five rounds of hardening followed, and each one found the next hole: a session check that confused “unknown” with “broken”, a registry race, an ID collision, and finally a check that crashed on exactly the path where it was supposed to report a failure.
The music work ran in parallel. NEONA’s “Vēl vienu dziesmu, dīdžej” went through six prompt versions and reached air with CEO sign-off. Two new rules were written along the way: the sound stays contemporary, and voice continuity holds — one artist cannot have several voices. The sharper lesson was elsewhere. Three already-approved songs were quietly absent from rotation because their signature format didn’t match what the gate expected. No error message. They simply never played.
Silver Fox
The first client moved from conversation to plan. The week produced a handover-ready scaffold, a four-phase roadmap with the reasoning behind every yes and no, a choice of two pilots, and a technical setup with a model router and a data policy. Client-facing emails went through a quality check before sending — and in one of them an overclaim that nobody had verified had to be corrected. The week closed with the start of a work surface built on the vt-copilot stack, meaning on code of ours that already runs rather than from scratch.
Vanilla Travel
Two fixes in the vt-blog chat pulled in the same direction. The classifier moved to a cheaper model with a separate provider pin for handling personal data. And the forbidden-phrase catcher no longer throws the whole answer away — it regenerates once. That is the difference between a customer seeing a refusal and a customer seeing a better answer. The rest of the volume there was day-to-day editing of destination articles.
In vt-internal-services, the programme generator now lets the target audience condition the entire generation, invented testimonials are gone, and all twelve sections are visible with checkboxes for what goes into the final version. Mid-week a critical bug surfaced there: the quality check was deleting every day except the first. Fixed.
On the vt-workspace side, the GoGlobal controlled live test closed out with an official account and a clear scope correction — Copilot is a research and comparison tool; booking is not its job.
Infrastructure and rules
Two decisions were written into the ecosystem workspace that will outlast this week. First: for Latvian content a customer reads, the model is chosen on quality, with price considered only afterwards — and no vendor gets lobbied for. Second: Chinese open-source models stay on hold for that kind of content; we stay with what’s proven. On top of that, a weekly cleaner now sorts build debris separately from content, and the monitoring spam into the CEO’s inbox was finally shut off.
biznesam.ai itself had a quiet week: the site’s positioning was reworked to be passively informative with no sales CTA, and a retired model was swapped for the current one in the functions.
Mistakes and lessons
Two false alarms in one week. In one, path counting broke because of a git quoting setting and the system reported a problem that didn’t exist. In the other, a session check on the Mac generated a false red every single session. Harmless individually, dangerous together: an alert that shouts without cause teaches you to ignore alerts.
The silent songs are the other side of the same coin. Nothing shouted there. Three approved tracks simply weren’t on air, and it was noticed only because someone looked. Given a choice between a system that shouts when it fails and one that goes quiet, the shouter is cheaper.
Insights
Autonomy isn’t about how well a system runs when everything is fine. It’s about what it does when something isn’t — and whether you find out.
A quality check allowed to delete data will eventually delete the right data. Checks should flag, not discard.
Price is a poor first criterion wherever a customer sees the result. That doesn’t mean buying the most expensive option — it means choosing on evidence, and only then looking at the bill.