Architecture

Control-plane placement

Qubanta sits between applications and model providers. It evaluates outputs as cognitive events, computes stability and confidence, then emits deterministic control actions suitable for automation and review routing.

Reliability model Developer contract

Inference path

The gateway pattern preserves existing application behavior while introducing a governed response surface: confidence bands, stability state, failure flags, and an action directive.

Client / Service
  → Qubanta /v1/infer
    → Model Provider(s)
    ← Raw Output
  → Qubanta Evaluators (constraints, stability, grounding, anomaly, health)
  → Confidence Engine (pressure + collapse rules)
  → Policy Engine (deterministic action)
  ← Governed Response (output + state + action)

Gateway

Request normalization

Input is normalized into a stable task envelope: task type, constraints, and optional policy overrides. This enables repeatable evaluation and consistent logging semantics across callers.

Signals

Evaluator pipeline

Evaluators produce bounded signals (0–1) for constraint integrity, stability, grounding strength, anomaly pressure, and system health. Signals are observable; no truth-probability claims are made.

Policy

Decision control output

Policy converts confidence state into action: accept, warn, review recommended, fallback suggested, or blocked. Downstream systems use actions to gate automation deterministically.