Dependent additive layers after an alarm¶
AdapterLoop(..., alarm_cascade="downstream") explicitly declares that later
additive layers depend on every upstream routing identity and regime. The
existing default, alarm_cascade="layer", retains layer-local recertification.
The downstream option requires additive mode.
A non-ignored alarm restarts the alarmed layer and every later layer, including their usefulness certificates, clears their old samples, resets downstream risk detectors, and starts their configured probation. An active candidate is discarded and its trial boundary removed before old-zone evidence can be reused. Retained layers keep their frozen routing boundaries. Existing rollback actions detach the requested suffix; detached models cannot return through additive rollback.
Restarts use the existing shared delta ledger and lazy charging. They do not lower risk levels, evidence floors, or promotion gates. Checkpoints retain this option, probation, and certificate state; older checkpoints default to layer-local behavior. Replay uses the same declared option. Ignored alarms remain ignored by policy.
This option responds to observed alarms. It does not detect an arbitrary hidden model swap or establish that an unchanged provider has an unchanged distribution.
Empty trial zones¶
DefaultPolicy(stale=...) already discards an uncertified trial after the stated
number of all decisions since freeze, including decisions outside its zone.
An empty zone therefore reaches a deterministic maximum age without relaxing the
certification or novelty sample requirements. The policy and decision clock persist
in checkpoints. Pause prevents new training; it does not suspend the trial deadline.
Discard removes only the candidate's trial routing boundary. Set the deadline when
constructing the policy; no second occupancy counter or minimum-evidence override
is needed to terminate a zero-occupancy trial.
This does not bypass escalation inside a nonempty trial zone. An alternative shadow label policy would need an explicit sampling and propensity ledger, separate validation of which labels may update each promise, and durable replay of that choice. Such a protocol is outside this option.