🤓 Latest Submissions
Tablekeeper Dark Factory
Tablekeeper Dark Factory is a software factory of four Claude Code seats in one Band Desktop room. A coordinator plans and routes but owns no code. An implementer builds the service. An auditor writes black-box probes from the requirements before any code exists. A reviewer rebuilds in a clean container with no network, runs the official harness, the auditor's probes and its own hand checks, then accepts or rejects. The factory completed all four Tablekeeper stages and passes every shipped check from a fresh clone. In stage 2 the reviewer rejected the first build for a browser defect the shipped samples never tested; the fix was accepted in round two. The human sent four dispatches and five wake-ups after provider stalls, none about the work. The mandates name no track detail, so the same seats can be pointed at another problem.
6 Oct 2026
Lexicon
Live captions break on the words a lecture is actually about: names and course-specific terms the recogniser has never seen. A student who can't hear the room has no way to tell a caption is wrong. With Lexicon, you paste the course glossary and start a session. Your microphone streams to AssemblyAI's real-time recogniser with those terms sent as keyterm boosts, and large, high-contrast captions appear with speaker labels. During the lecture you can type a question. The answer comes only from what has been captioned so far, with the exact caption lines it relied on. Click the citation to jump back to that moment. The server checks every quoted passage word for word against the captions. If the check fails, or the captions don't cover it, you get "not found", never an answer from general knowledge. When you press Stop, you get a summary and key terms defined as this lecture used them, each tied to caption lines, plus a report of glossary terms that never appeared as written. Everything exports as Markdown. Measured, not claimed: on a recording full of unfamiliar surnames, 4 of 10 terms appeared as supplied only with boosting ("n-k Dirac" became "Nkemdirim"). On standard linear-algebra vocabulary boosting made no difference, because the recogniser already got those words right. A 20-minute test run finished with no crash and no reconnects. Stack: React client, Hono proxy that keeps both API keys off the browser, AssemblyAI streaming, OpenAI Structured Outputs. No database: everything is session-scoped.
30 Sep 2026
