Loading...
Loading...
OmegaEngine · Arbitration
Multi-Model Arbitration sends the same prompt to multiple LLMs simultaneously, scores each response against your policies, and returns the one with the lowest risk — automatically.
Your prompt is sent to 2–6 models in parallel via the Gateway. Responses stream back concurrently.
Each response is evaluated by the policy engine: risk score, safety flags, compliance tags, and cost.
The response with the best composite score is returned. Losing responses are logged for audit.
Every arbitration event records per-model signals so you can see exactly why the consensus won:
| Signal | Description |
|---|---|
| consensusWins / consensusWinRate | How often each model's answer matched the selected consensus |
| avgDistanceFromConsensus | How far a model's responses sit from the consensus verdict |
| avgLatencyMs | Average response latency per model |
| avgTotalTokens / avgCostUsd | Token and cost telemetry per model |
Arbitration runs inside the decision pipeline; the observability endpoints below expose the events and per-model insights (Business plan and up), always scoped to your own tenant:
# Recent arbitration events — full model responses + why the consensus won
curl -H "x-api-key: $OMEGA_API_KEY" \
https://omegaengine.ai/api/arbitration/events
# Per-model insights over the recent window
curl -H "x-api-key: $OMEGA_API_KEY" \
https://omegaengine.ai/api/arbitration/insights
# → {
# "windowHours": 24,
# "byModel": [
# { "modelId": "claude-opus-4-8", "provider": "anthropic",
# "calls": 412, "consensusWinRate": 0.71,
# "avgLatencyMs": 1840, "avgCostUsd": 0.004,
# "avgDistanceFromConsensus": 0.09 },
# ...
# ]
# }
# Watch it live: /dashboard/arbitration (streaming view + replay + intervene)Compare model outputs for investment advice, loan decisions, and fraud alerts — pick the most conservative.
Validate clinical suggestions across multiple models to reduce hallucination risk.
Cross-check contract analysis and regulatory interpretation for accuracy.
Let your agent query multiple models before taking irreversible actions.