universal_reflex_engine.md

Universal Reflex Engine (URE)

Overview

The Universal Reflex Engine is IEE v3.0's real-time, structure-driven intelligence layer.

URE operates without per-step LLM calls and without heavy training loops in core runtime.

Processing Loop

observe -> extract -> model -> afford -> decide -> (optional) act

Core Contracts

UniversalFeature

• type: domain-agnostic classification
• bbox: spatial footprint
• actionable: execution candidate flag
• dynamic: temporal movement flag
• salience: relative importance score

WorldModel

• tracked objects
• relationships (proximity, overlap, motion, hierarchy)
• frame signature + change count

Affordance

• object-to-actions mapping
• deterministic action primitives

ReflexDecision

• selected object + action + reason + priority
• executable flag
• exploratory flag

ExperienceEntry

• compact world-state summary
• executed decision
• reward
• timestamp

Safety and Execution

URE does not bypass core execution contracts.

• Optional execution path goes through ActionExecutor.
• Policy gate (PermissionPolicyStore) is checked before execution.
• Exploration remains bounded and policy-aware.

API Surfaces

• GET /ure/world-model
• GET /ure/affordances
• GET /ure/decision
• GET /ure/metrics
• GET /ure/experience
• POST /ure/step
• POST /ure/demo

Performance Signals

Per step:

• decision_time_us
• loop_time_us
• decision_within_budget

Aggregates:

• average decision ms
• p95 decision ms
• average loop ms
• over-budget count
• exploratory decision count

Adaptation Model

URE adapts through interaction outcomes, not offline training.

• successes reduce local failure bias
• failures increase local failure bias
• affordance confidence is adjusted by outcome

This keeps adaptation fast, bounded, and deterministic.