Execute

The vault has a resident operator. I call it the Butler - an unattended agent that wakes every thirty minutes, checks whether anything is pending, and if so runs one working cycle on my behalf. A gate script keeps empty ticks free, and when it does wake properly it runs on a cheaper model with a scoped tool allowlist that can't send, publish or spend. Everything around it is deterministic, in that code decides whether the Butler wakes, what it may touch and what gets logged, and the model is only the judgement in the middle.

One cycle covers the whole machine. The Butler sweeps any raw voice captures into finished notes, reads and acts on my answers to its open questions, classifies new tasks, runs the closure pass over anything that looks finished, checks every pipeline lane for failures, and writes up what it did.

Every task it touches gets classified, and ambiguity always demotes. A task it can do end to end is AUTO. A task where it can produce something useful for me to finish is PREPPED. Everything else is HERS and stays untouched. The tests are concrete (could this be done from the command line, and would a prepared artefact actually save me time?), and when the answer is unclear the task moves down a tier, never up.

Hard gates override everything. It never sends anything, never deletes without my word, never publishes externally, never spends, and never touches the fields and sections that are mine alone. The content, email and transcript lanes have their own judgement loops, and the Butler stays out of them entirely.

The Butler can repair the machine, within strict limits. When a pipeline lane breaks and the fix doesn't need me, it makes the fix itself - config, thresholds and templates only, backed up first, checked against everything that depends on them, and never twice for the same failure. The carve-out is a precise one, in that it may patch a lane's script but never its output. When the fix does need me, the Butler asks instead.

Everything lands on two surfaces. Needs You is the single place anything wrong or waiting for me appears, anywhere in the system - its questions, prepared work ready for review, pipeline failures, unfiled counts. Claude Did is the audit trail, one block per cycle, and a Sunday job asks me to skim it weekly and calibrate it.

The trust is earned, and checked. The Butler acts, the audit trail records it, I calibrate it weekly, and its rules absorb the corrections. I delegate to it, but with a paper trail rather than on faith.