The interesting work was rarely the feature. It was the second-order problem the feature
exposed — and in more than one of these the obvious fix was tried first and turned out to be a regression.
The feed was crediting a noticeboard as a publisher
Symptom. The news surface promoted a personal newsletter alongside Google Research, and dozens of stories carried the byline “Hacker News”.
What was actually wrong. Hacker News is a discovery channel, not a publication. Every story arriving through it lost its real provenance. Rather than assume the scale of it, we sampled 100 links and found 72 distinct domains behind them.
Fix. A domain-based re-attribution layer against a map of 34 known publishers. Every source now carries a kind — primary, press or community — and that axis governs ranking and placement, not just a badge. Personal blogs were kept but barred from the lead slot: the objection was to promotion, not to inclusion.
Evidence. Zero aggregator bylines remain. Current mix: 39 primary, 15 press, 6 community.
The lead slot went blank exactly when the news was biggest
Symptom. The hero carousel rendered an empty column on some slides.
What was actually wrong. Most lab announcements ship without a preview image. In one run, none of Google AI, OpenAI or Anthropic had a usable thumbnail. The obvious fix — prefer illustrated stories — was tried, and it was a regression: it dropped OpenAI and Anthropic out of the rotation on precisely the days they shipped something. It won a picture by losing the point of the section.
Fix. Image preference now applies within a source, never between them. When a lab publishes without a picture the frame becomes a masthead, the organisation’s name set large. On a story from a frontier lab, whose desk it came from is the strongest thing the slot has to say, so it gets said at full size.
Evidence. The rotation control is a ledger of source names rather than a row of dots — dots report how many slides exist, which is the least interesting fact available.
A job that finished in 19 seconds ran for 20 minutes
Symptom. The news refresh workflow was cancelled at the 20-minute ceiling, despite logging completion.
What was actually wrong. The work finished in 19 seconds; the process then refused to exit. Rather than guess, we measured the active handles and found two independent holders of the event loop: the database channel, and five keep-alive sockets left over from feed fetching. An earlier diagnostic run had produced a misleading reading, so it was discarded and the test rerun properly before any fix was written.
Fix. An explicit shutdown that terminates the client and deletes the app, plus an unreferenced fallback timer that only fires if something else is still holding the loop.
Evidence. 1,233s cancelled → 49s, 50s and 53s success across three consecutive runs.
The interface had no elevation, so it faked it
Symptom. The UI read as templated, and nobody could say precisely why.
What was actually wrong. The canvas and the cards were the same colour. With no tonal step between the page and the things sitting on it, every panel had to be separated by shadow and glow. Two duplicate colour families were also in play, running alongside the brand neutrals and the brand green.
Fix. A real canvas-to-card step, the duplicate families deleted, and structure carried by hairlines instead of shadows. Theme tokens are exposed as CSS variables, so one block governs both themes. Charts were the awkward case, because chart colours are values rather than classes; a theme hook watching the root element solved it.
Evidence. An honest finding is recorded in the code: no single green clears the 4.5:1 contrast threshold on both the light and the dark ground, so the two themes take different tints deliberately.
One model retirement took out six features at once
Symptom. Generative features began returning 404s.
What was actually wrong. A preview model had been retired. It was named at six call sites plus the digest default.
Fix. Two named constants — a reasoning model and a fast model — behind a wrapper that detects unavailability and falls back automatically. All eight call sites now route through it.
Evidence. It paid for itself immediately: the key in use turned out to be free-tier, so every request to the larger model returns a quota error. Instead of six broken features, the platform runs on the fast model — and that was surfaced to the operator as a finding rather than hidden by the fallback.
