Machine Learning
Agent Orchestration Lab
Wire cheap→expensive model cascades with deterministic escalation gates and cost budgets.
Agent Orchestration Lab
FDE · cost-aware cascadeCheap model scores every case. Deterministic gates (not the model) decide who gets the expensive model. Tune gates to catch hard cases under budget.
| Case | Cheap score | Unc. | Stakes | Gate | Needed expensive? |
|---|---|---|---|---|---|
| Routine FAQ paraphrase | 0.92 | 0.10 | low | keep cheap | no |
| Borderline policy edge case | 0.61 | 0.55 | high | escalate | yes |
| High-confidence wrong pattern | 0.88 | 0.20 | med | keep cheap | yes |
| Obvious spam / out of scope | 0.15 | 0.05 | low | keep cheap | no |
| Legal-adjacent claim | 0.70 | 0.45 | high | escalate | yes |
| Simple classification | 0.95 | 0.08 | low | keep cheap | no |
| Multi-hop research brief | 0.55 | 0.60 | med | escalate | yes |
| Customer-critical escalation | 0.40 | 0.70 | high | escalate | yes |
Total cost
56 u
Escalations
4
Recall (catch hard)
80%
Precision
100%
Missed hard cases (false negatives): 1 · Wasted escalations: 0