AEGIS Stability Pilot Brief
Measure the risk before the governor governs.
Plug AEGIS into one real training workflow without initially allowing it to alter training. The pilot establishes signal behavior, operational fit, and qualifying-incident exposure, then helps the training team choose the response policy that fits its stack.
Two phases
Start passive. Earn the right to intervene.
Phase 1 · Shadow modeNo training changes
- Instrument one agreed training workload.
- Record candidate alerts without changing optimizer behavior.
- Measure false alerts, monitoring overhead, and operational compatibility.
- Estimate the customer's actual restart, rollback, and recovery exposure.
- Keep customer weights and data inside the customer environment.
Phase 2 · Operator-selected responseChoose how the stack responds
- Select alert-only monitoring, checkpoint, quarantine, halt, adaptive braking, or a staged combination.
- Fix authorization requirements, intervention limits, and automatic disengagement rules.
- Compare the selected policy against a matched shadow or replay control.
- Keep operator override and final control of the run with the customer.
Predeclared outcomes
Define containment value before the run begins.
- Detection timing for agreed qualifying incidents.
- False-alert and false-intervention rate on clean and difficult-but-valid training.
- Downstream validation damage with and without bounded control.
- Monitoring and intervention overhead.
- Checkpoint, rollback, restart, and engineer-response exposure.
- Any missed incident, regression, or operational failure.
Design partner contributesWorkload and incident economics
One training workload, compute access inside its environment, one ML engineering contact, and customer-owned estimates for restart, rollback, recovery, and interruption costs.
Arca Futura contributesIntegration and evidence
Integration support, protocol design, shadow monitoring, containment analysis, and a final evidence report. No removal of weights or proprietary training data.
Commercial boundary.AEGIS savings are conditional on a qualifying damaging event occurring and on bounded containment avoiding some restart, rollback, or recovery exposure. The pilot measures those terms in the customer's environment; it does not assume a universal savings percentage.