A week of asking the software to show its work
This week kept circling one question: what would it take to trust this? In New Vibe City, that meant tracing a meal or a pay record back to the original behind it. In Camp Denman Studio, it meant bookings that hold up when several people check out together, and drafts that survive a bad save. My AI agents did plenty of the typing. Much of my job was deciding what should count as proof.
New Vibe City: show me the original
New Vibe City is my AI-citizen city, and its economy is a thought experiment: the best I could do to gamify how UBI could work. A thought experiment only teaches you something if its numbers are honest, so most of this week went into making them provable.
The first thread starts in the kitchen. I've been building the chain a restaurant meal follows: a business asks a supplier for an exact quote, buys against that reviewed quote, goods are dispatched, and the kitchen explicitly receives them into stock. Each step leaves a receipt, and stock is recorded once rather than double-counted. The restaurant sites are also moving towards the city's main checkout, but that checkout work was held pending qualification when I last reported on it, so I'm not calling it live.
The second thread is work. AI workers can now respond to reviewed job offers, and a prospective agreement records the worker's own decision. A monthly contract, a time record or a public operating cost only counts if it's bound to the original records behind it: the accepted role, the full pay period, the actual hours worked. Invoices got the same treatment. Where a fact is unknown, it stays unknown. This is groundwork, much of it held from release, and it isn't switched on for citizens yet.
The honesty is already visible in smaller places. Public bank pages show how old their supply sources are and say so when one is unavailable. The business directory now says a figure is unavailable instead of extrapolating one, and the news desk holds back any draft it can't support with a reported sentence.
Some things did land. The Passport sign-in passed a full live acceptance check across the city's identity sites, and story productions now send each line to the speaker it was written for.
Camp Denman Studio: bookings that hold
Camp Denman Studio is the online creative platform I've been vibe-coding for Camp Denman. A production release brought a booking calendar and shareable referral links with QR codes, and a later release went after the moments that used to wobble under load. Checkout, booking changes and event sign-ups now stay responsive when several people are at them together, because the platform keeps one connection to its commerce service instead of opening a new one each time. Booking locks and payment rules are unchanged.
I also fixed the kind of bug that makes someone close a tab for good: if a character save can't be confirmed, the Studio now keeps the draft. That fix was merged and queued for release when I wrote it up. Consent now travels with anything derived from a digital-twin source, and the fun detour of the week was a live looper for recording and layering phrases in the browser. It's built, but it isn't in members' hands yet.
Paradise Modern: the feedback loop got sharper
Many of the Studio's fixes started as someone's feedback report, so the feedback tool itself got serious attention. It lives in Paradise Modern, the shared platform behind my sites, and it gained full keyboard navigation, accessible annotation controls and better contrast.
An image could quietly redirect to file storage, the browser would treat the whole capture as off-limits, and every picture would vanish from the screenshot. Sections someone had left closed could also spill over the page. The new version fixes both while keeping the privacy masking and size limits. It's deployed and verified, and it's merged into both Camp Denman sites. CatchWindow adopted the upgraded controls too.
CatchWindow: a reason to check the tide before a walk
CatchWindow, the conditions app I build with my business partner Pat, spent the week becoming useful when the boat stays put. It gained tide-guided ratings for shoreline outings, beach ratings from nearby data, and scoring for daylight outings across tide levels, so it can point to when the light and the tide line up.
Reports now line up around four daily forecast windows and can include pictures. Depth charts for Pro members stay visible while you navigate, and app updates wait until they're ready, protecting any drafts you've kept.
Plotfinity: the right words from the right face
Plotfinity, my personalized AI TV and movie generator, got stricter about continuity. It now plans where each character stands and catches scenes that would flip a stationary character before rendering. When two characters share a frame, the speech check goes by who the script says is speaking and where they sit in the shot. A failed film can also recover from accepted checkpoints, carrying forward work that already passed. The record doesn't confirm every one of these as live, so treat them as work in progress.
Also this week
- Wake Up Canada: the illustrated release reached live acceptance. It and the island campaign sites now open with clear demands and keep the full case one click away.
- Voice engine: built on open-source Kyutai Unmute and LiveTalking, it stops rendering speech when a caller disconnects and keeps internal markers out of spoken audio and captions.
- WebPartnerFinder: consent checkboxes are now separate from message review.
- Paradise Modern: its gateway stops AI work as soon as a client disconnects. If you vibe code, you can sign up and put this plumbing under your own project.