MULTI ENGINE BROKER

Multi-Engine Thought Broker¶

Mobius Broker operates as a constitutional router that can evaluate a prompt across several engines (Gemini Antigravity, OpenAI, Claude, DeepSeek, and local Sentinels) while preserving the AI→AI break by routing all final decisions through Sentinel consensus and the Civic Ledger.

See Also: For the comprehensive Multi-Engine Model Taxonomy (MEMT) that classifies engines by cognitive specialization and provides intelligent routing, see MEMT Documentation.

Routing Modes¶

The default routing mode is local, but callers can opt into other modes via the routingMode field on /v1/deliberate.

Engine Registry¶

Each engine has a typed configuration (EngineConfig) that captures id, baseUrl, apiKeyEnv, defaultTools, and maxTokens. This registry enables feature flags (e.g., ANTIGRAVITY_ENABLED) and keeps the router aware of which engines are safe to target in production.

Deliberation Flow¶

1. Prompt submission — a human or downstream app calls POST /v1/deliberate with a prompt plus options (routingMode, engines, allowedTools, etc.).
2. Engine fan-out — the router selects engines based on the routing mode or explicit engines array and invokes each engine client (antigravityClient, openaiClient, claudeClient, deepseekClient). Each invocation captures latency, tool traces, and risk flags and emits an external trace for audit.
3. Sentinel evaluation — regardless of external engines, the router always calls the Sentinel consensus service (if configured) or falls back to local Sentinels. This step yields final_answer, gi_score, decision, and flags.
4. Ledger attestation — the router submits the deliberation metadata to LEDGER_ATTEST_URL, recording GI scores, routing decisions, and per-engine risk telemetry.
5. Human-gate — if gi_score < 0.95 or the consensus decision asks for human review, the API returns needs_human_review (HTTP 202) and surfaces the flags to operators instead of emitting an automated answer.
6. Response — on success the caller receives { gi_score, answer, engines, deliberation_id }. Each engine result contains engineId, answer, riskFlags, and latencyMs, making it easy to visualize how the Sentinels weighted each candidate.

Guardrails¶

• No engine result is returned directly to users; Sentinels must evaluate every candidate.
• Every deliberation produces a Civic Ledger attestation to maintain an auditable chain of custody.
• Remote consensus failures fall back to local Sentinels only if requiredSentinels are provided, preventing silent degradation.
• External traces are emitted per engine (provider, riskFlags, latency) to support retrospective investigations.

MEMT Integration¶

The broker now supports MEMT (Multi-Engine Model Taxonomy) routing via /v1/memt/deliberate. MEMT provides:

• Task Classification: Automatically classify tasks as ARCHITECTURE, ENGINEERING, FRONTEND, MATH_OPTIMIZATION, MEMORY_RECALL, CIVIC_POLICY, CRITICAL_DECISION, or GENERAL
• Cognitive Routing: Route tasks to the most appropriate engines based on their cognitive specialization:
• ACI (GPT): Architect-Class Intelligence for planning and synthesis
• ENI (Claude): Engineer-Class Intelligence for code and verification
• SXI (Gemini): Software Operator Intelligence for UI and tools
• OEI (DeepSeek): Optimization Engine Intelligence for math and performance
• MSI (ECHO): Memory-State Intelligence for caching and recall
• DVA Tier Support: Automatic tier inference (LITE, ONE, FULL, HIVE) based on task risk and kind
• Dynamic GI Thresholds: Risk-based GI thresholds from 0.90 (LOW) to 0.98 (CRITICAL)

MEMT Endpoints¶

See MEMT Whitepaper for full documentation.