An L2 step (automation + approval) can be promoted to L1 (autonomous), but only with evidence, and only while the platform keeps watching it.
promotion evidence in the workflow definition (a configuration change), and no step ships pre-promoted. The drift machinery below is implemented but only applies once a step has been promoted.Promotion requires evidence that the step is reliable:
The promotion record captures the prompt, model, and served model at the time, so "promoted on model X" is auditable.
A promoted step isn't left unwatched. A fraction of its L1 outputs (its spot-check rate) are routed back for after-the-fact human review. Spot-check acceptance is tracked and feeds back into the step's health.
spot_check_rate field on the step's JobTemplate, and it defaults to 0.0 — promotion evidence alone does not turn sampling on. A promoted step must set spot_check_rate > 0 (enforced at definition time), otherwise no spot checks are created and the queue's Spot checks tab stays empty.spot_check_rate, so live runs generate no spot checks at all. The only exception is seeded demo data: the demo seeder fabricates spot checks at its own 20% rate (seed_demo.py), independent of the workflow definitions, so a seeded database can populate the tab.The platform re-checks whether a promotion still holds. A step is flagged as drifted if:
A drifted step falls back to L2: it keeps running, but a human must approve it again until the drift is resolved.