Skip to main content

Trust levels

A level describes the working agreement for one class of work. It is not a permanent rank for a team. Routine documentation may run unattended while an authentication change on the same board remains fully supervised.

LevelWho runs the work?When does a person step in?Entry point
L1The engineerThroughout/faff-wtf, /faff-map, /faff-jot, /faff-plot, /faff-prep
L2The agent, one change at a timeAt each major gate/faff-graft
L3The agent, across an eligible queueWhen work is ambiguous, blocked, or ineligible/faff-beep-boop
L4The agent, without scheduled supervisionWhen a guardrail refuses the run or evidence requires judgementfaff lights-out

L1: you run the loop

You do the engineering. SuperDomestique helps understand the backlog, choose work, and prepare a specification. No delivery authority has been handed to an unattended process.

L2: one supervised build

/faff-graft takes one specified issue through implementation and review. The agent does the work, but the important transitions still wait for approval. This is the normal place to learn how the system behaves on a repository.

L3: an unattended queue

/faff-beep-boop drains work that a person has made eligible. The run parks anything it cannot safely call, records why, and leaves the tracker as the morning view. The run ledger refuses a clean finish while admitted work remains unaccounted for.

L3 is the current unattended path. See Unattended runs for eligibility, parking, and run records.

L4: lights-out preview

faff lights-out adds a fail-closed preflight, budget and liveness controls, adversarial review, a holdout verdict, and merge interlocks. The machinery exists, but the complete L4 claim is not yet supported by external proof.

The distinction matters: invoking a guardrail can be mechanically enforced, while the quality of a model's review remains a judgement. The evidence page links the current audit and open proof work.

Choose a level by workload

Start with a narrow class of low-risk, well-tested work. Move it from L1 to L2, then to L3 only after the results justify less scheduled attention. Other work can stay at L1 indefinitely. The adoption guide shows how eligibility and appetite support that approach.