Safety and operations
Section

Safety and operations

What makes it safe to hand an AI organization real authority: one send gate, QA against user stories, gated self-repair, watchdogs, and durable state.

6 articles38 min total

Start reading →

Guardrails

  1. One door out: the send gateThe system can draft anything and send nothing without a yes. One chokepoint, content-bound approval tokens, a freeze switch, and a scan for bypasses.6 min read
  2. QA: user stories, Playwright and synthetic usersQA driven by user stories: browser journeys, accessibility, visual baselines, synthetic users, and a persona tester that never sends.7 min read
  3. Auto bug-squash and the WardenGated self-repair: fixes in isolated worktrees with red-to-green proof, and a governor that admits autonomy by reversibility, not confidence.6 min read
  4. The system documents itselfRender from what exists, not from a hand-kept list. The org chart lesson, the daily capabilities page, and the counting rule.5 min read

Keeping it running

  1. Watchdogs: liveness is work completedCatch-up after reboots, stale-job alerts, death-cluster detection, auto-unwedging, and the memory breaker that ended a kernel-panic loop.8 min read
  2. Durable state: Postgres as the system of recordIf it matters, it lives in a table, not a context window. Backups, the incident that tested them, and the move to local hardware.6 min read