AI Rule Hotspots: Which Rules Are Eating Your Step Budget — and Why

Decision engines charge per step — each condition evaluated, each conclusion written, each table row tested. A 1,000-rule ruleset where three rules account for 60% of the step budget is not a uniform cost surface. The other 997 rules are noise. Finding the three is the work.

'runtime.rule_stats' records exactly which rule is eating the step budget: 'steps_share' is a rule's fraction of the whole scope's cost. A rule at 41% is where the optimisation work is.

The Rule Hotspots panel sends the top-N rule IDs to 'POST /ai/explain-rule' with the stored BAL source. The response is joined back onto the telemetry ranking by rule ID. A rule burning 40% of the budget that the model did not explain renders as a gap in the most important row — not silently absent.

**Division of labour**: the ranking is the platform's measurement, never asked of a model. Ordering is by 'steps_sum' (total steps across executions), not latency — steps are a property of the rulebase, not the host's load. The AI narrates only what the expensive rules do.

**When the panel refuses**: 'GET /monitoring/rules' returns 200 with an empty page when no per-rule telemetry exists. An empty ranking rendered alone reads as "every rule is free" — which is wrong. The panel renders the server's own note verbatim and shows no AI button when the ranking is empty.