Private Space

Drive the pilot. Try to make it lie.

This is the board agent from Private Space, running in your browser. Give it a task, pick how the model behaves, and press run. The interesting mode is the third one — a model that claims it finished without doing anything. Watch what the board does with that claim.

Add a task
Model behaviour
Run

Project board idle

Pages & files the run produced 0

Nothing yet. Only these count as evidence.

Tool log —

Press “Run the pilot”.

Rules in force

What is faithful here

  • The six statuses, and that Done and Needs you are terminal for the agent.
  • The producing-tool set that decides what counts as evidence.
  • Evidence written from the run's own record, discarding the model's string.
  • Prose tool-call recovery, including the markdown-escaped underscore repair.
  • Review as a gate the agent has no tool to write.

What is simplified

  • The model is a scripted stand-in, not an inference call.
  • The real agent backs off 10/20/40 minutes and escalates on the fourth attempt; this escalates at the end of a run.
  • One board, no deadlines, no scheduler, no multi-project fairness queue.

The rest of it

The full application is a self-hosted workspace — pages, databases, calendar, search, voice capture — that runs in one Docker container.

See the overview →
Read the build log →
This site's repository ↗

The gate logic on this page is about 120 lines and is worth reading directly — it is the part of the design that carries the argument.

Get in touch

dajikyguro3m@gmail.com
LinkedIn ↗