A hybrid routing agent that answers every task locally at zero counted tokens first, and escalates to Fireworks AI only when the local answer can't be trusted — across all 8 required task categories.
Verification is deterministic per category — a locally-generated Python expression cross-checks math, an AST parse validates code, entity-union catches missing NER labels, a label-consistency vote guards sentiment. Escalation only fires when local verification cannot confirm a safe answer, so most tasks never touch a counted token.
Internal benchmark against our own judge model — not a leaderboard score. Measured on a 2 vCPU / 4 GB simulated judging environment.
| Category | Local safeguard |
|---|---|
| Factual Knowledge | Complete-answer hint, two-part question detection |
| Math Reasoning | Local counter-calculation (Python expression, cross-checked) |
| Sentiment | Label-majority vote across local runs |
| Summarisation | Format/length gate + local retry on violation |
| Named Entity Recognition | Entity union + name/date normalization |
| Code Debugging | AST syntax check + identical-code rejection |
| Logic Puzzles | Escalated by policy (local model unreliable here) |
| Code Generation | AST syntax check + local retry on prose-only answers |