What it is vs what it is not. Written after external review.
A reactive control system with decision caching.
Three layers run in priority order:
- Reflex: hardcoded rules, fires instantly
- Cache: replays past LLM decisions for similar states
- Brain: LLM fallback for unknown states
The cache reduces LLM calls over time. Cost approaches zero. Latency approaches zero. This is real and measured across 5 adapters.
- Not a learning system. It caches, it does not learn from outcomes.
- Not an AI agent. It does not plan, reason, or generalize.
- Not a neural network. Pattern store is a hash map with fuzzy matching.
- Not a replacement for RL. There is no reward signal.
- Not safe by default. A wrong LLM decision gets cached and replayed.
What we say: "Brain teaches once. Muscle remembers forever." What is true: Brain decides once. Cache replays that decision blindly. What is missing: No evaluation of whether the decision was good. What would fix it: Outcome scoring after every action. Reinforce good decisions, suppress bad ones. Planned for v0.2.0.
What we say: Developer defines feature extractor. What is true: The developer decides what features matter. If they pick wrong features, the cache is useless. What is missing: No automatic feature discovery. No spatial awareness. No temporal context. What would fix it: Richer state representation. Learned features from raw sensor data. Planned for v0.3.0.
What we say: "Real-time AI execution runtime." What is true: Each frame is independent. No memory of what happened 1 second ago. What is missing: Trajectory awareness. Sequence decisions. Planning over time. What would fix it: Sliding window of recent states as context. Planned for v0.3.0.
What we say: Reflex layer handles safety. What is true: Reflex handles hardcoded rules (stop if too close). But LLM decisions go straight to cache without validation. What is missing: A gate between LLM output and cache storage. Transient vs verified tiers. What would fix it: L1 (transient) + L2 (verified) cache tiers. Safety check before promotion. Planned for v0.2.0.
What we say: "Muscle handles 99%. Brain handles 1%." What is true: The muscle replays decisions. It never checks if they worked. What is missing: A reward signal. Score after every action. What would fix it: Evaluate state_after vs state_before. Reinforce or suppress. Planned for v0.2.0.
What we say: 10% exploration rate. What is true: 10% of cache hits are randomly sent to brain for verification. This catches drift. What is missing: No epsilon-greedy over alternative actions. Only verifies the same decision. What would fix it: True exploration: try different actions, compare outcomes. Planned for v0.3.0.
These are not oversold. They work as described.
- Cost reduction is real. Brain calls drop from 27/min to 0/min. Measured across 5 adapters.
- Cross-body works. Same protocol, same brain, 5 different bodies. Zero code changes.
- Cross-language works. JS + Python via HTTPTransport. MuJoCo proven.
- Real physics works. Cart-pole balanced 64+ loops with 89% cache rate.
- Three-layer priority works. Reflex fires before cache. Cache fires before brain. Order is correct.
- Similarity matching works. Fuzzy feature distance finds close-enough patterns.
- Exploration catches drift. 10% verification rate detected brain drift in testing.
v0.1.1 (now): Reactive caching. Works but does not learn.
v0.2.0 (next): Confidence gating. Safety validation. Outcome evaluation.
v0.3.0: Small classifier. Temporal memory. Actual learning.
v0.4.0: Live distillation. Model trains during operation.
v1.0.0: Stable API. Learned model proven. Safety certified.
Each version fixes a specific weakness. No version claims to be more than it is.