Canary Rollout of Control Policies
A canary policy gains narrow, bounded authority — small setpoint deltas inside tight envelopes — that widens only as monitored performance holds, with instant revert on breach.
Authority in small, reversible increments
A controller that survives shadow has proven it would not obviously misbehave; it has not proven the machine responds as the model expects when it actually acts. Canary bridges that gap by granting minimal authority: the new policy may command only small deviations from the incumbent's setpoints, inside envelopes far tighter than the machine's true limits, and only during nominal operation.
The canary's authority is a scalar that ramps. It begins near zero, and each monitoring window that passes without a breach widens the allowed deviation and the operating regimes it covers. Any breach collapses the scalar to zero instantly, handing control back to the certified incumbent. This is a continuous, reversible handover, not a switch.
Canary guardrails
- Bounded command delta blended with the incumbent output
- Restricted to a subset of the validated envelope
- Tight rate limits and hard clamps at the L1 edge
- Armed rollback to the incumbent at all times
- Human oversight for each authority-widening step
def canary_command(state, incumbent, candidate, alpha, env):
u_i = incumbent.act(state)
u_c = candidate.act(state)
u = u_i + alpha * clip(u_c - u_i, -env.max_delta, env.max_delta)
if not env.contains(state) or candidate.confidence(state) < env.c_min:
return u_i # candidate mute outside its regime
return u # alpha in [0,1], ramped by monitors
On the breeder, canary might first touch only slow shape corrections around delta -0.30, never disruption-critical control. On the burner, given the honest gates, canary authority over plug and DEC systems stays extremely narrow. Widening to full PROD requires the approval quorum. This is the machine-facing analogue of a progressive rollout, adapted to hardware where a bad step is not a bad user session but a fault.