Reliability Engineering
Almgren-Chriss Scheduling for Compute-Constrained Reasoning
An optimal-execution framework adapted to pace reasoning workloads and remediation rollouts under finite compute and action budgets.
Mohan Murari, Raghav Balasubramaniam — Raincurve Research
Abstract
Continuous reasoning competes for a finite compute and action budget. Raincurve adapts the Almgren-Chriss optimal-execution framework to schedule reasoning workloads and remediation actions so that aggressive execution does not destabilize the systems being monitored and conservative pacing does not delay resolution unnecessarily.
Scheduling model
The framework treats compute allocation and remediation rollout as a tradeoff between immediate system and operator load on one hand and delayed-detection or delayed-remediation risk on the other. A tunable risk-aversion parameter lets each environment choose the right execution curve.
Applications
The scheduler governs whether to batch pre-filter scoring or process trajectories as they arrive, and whether a verified remediation should roll out across an affected footprint at once or in carefully staged increments.
Conclusion
By making timing and impact explicit rather than relying on fixed heuristics, Raincurve can reason continuously without becoming a new source of operational instability.