What should a trading runtime process first during a trigger storm?

During a trigger storm, the runtime should preserve the deadlines attached to existing exposure before spending scarce capacity on every new opportunity. We propose testing that scheduling hypothesis with a saved burst of wakeups and a fixed service budget. The result remains a design question, rather than a measured improvement or a current DXAP scheduling claim.

The continuous record companion describes scheduled turns alongside agent-created triggers. Its reported protection failures make deadline handling worth examining, although trigger capacity and inference queues are separate mechanisms. The controls paper companion anchors the distinction between model decisions and the execution state they must manage.

Forty seconds of capacity

An illustrative burst contains sixty tasks arriving together. Ten reconcile open positions and have a twenty-second deadline. Fifty investigate new entries with a two-minute expiry. Four workers each need eight seconds per task. The first forty seconds provide twenty completions: four at each of five completion times. A queue that places all fifty research tasks first completes none of the ten reconciliations within twenty seconds. A policy that starts reconciliation work first completes eight by sixteen seconds and the remaining two by twenty-four seconds, leaving two deadline misses even under its preferred order.

That arithmetic identifies both prioritization value and insufficient capacity. It prevents a favorable scheduler comparison from implying that the resource envelope meets every obligation. Add varied service times, worker failure and stale-state refresh to the next fixture. Declare whether a task deadline applies to starting inference, finishing inference or obtaining reconciled state; those are different endpoints.

Priority should come from authenticated task type and current exposure, rather than free-form urgency written by a model. A low-confidence risk alert also needs a route to verification so false alarms cannot monopolize the queue. Deduplicate wakeups tied to the same condition while preserving a receipt for each suppressed task.

What would count as a useful result

Compare first-in-first-out, exposure-first and a deadline-aware policy under the identical arrival trace. Score missed risk deadlines, elapsed time with unresolved exposure, completed opportunity reviews and computation wasted on expired tasks. Carry dropped work into the report with its reason. A queue can reduce one failure class while increasing another, and the tradeoff should be visible.

We would hold the action policy fixed so any difference follows from scheduling and information age. All intended orders stay inside replay. The research output would be a capacity curve showing how many obligations each scheduler can complete before their declared deadline. If priority only helps at moderate bursts, that boundary guides infrastructure sizing. If it starves routine reconciliation after repeated alarms, the next repair belongs in admission control or priority aging, with its own fixture.