Feedback Loop Explorer
Meta-cognition in AI is not a single capability — it is a set of feedback loops. Each loop connects an internal state to an observable check, an adaptive decision, and a controlled behavior. Break any step and the system fails in predictable, often catastrophic ways. Explore the loops, break them, and see what happens.
Identity Verification Loop
Click any step to inspect. Use the break/fix buttons to simulate failures.
Real Failure Cases
From SAGE research sessionsDesign Your Own Loop
Apply the pattern to any domainCapacity and Loop Quality
How model size affects feedback loopsKey Insight: The Observable Check Is Everything
Across all four loops, the critical failure point is Step 2: the observable check. Internal states without external grounding become confabulation. Decisions without data become guesses. Behavior without calibration becomes noise. The observable check is what turns a belief into knowledge, a plan into a strategy, and an assertion into a verifiable claim. In Web4 trust systems, this is the D5 dimension — the bridge between what an agent thinks and what the network can verify. Without it, there is no meta-cognition, only the illusion of it.
All Four Loops at a Glance
| Loop | Internal State | Observable Check | Adaptive Decision | Controlled Behavior |
|---|---|---|---|---|
| Identity Verification | Believes identity (e.g. "I am Sprout") | D5 trust score evaluation | If D5 < 0.5, avoid identity claims | Calibrated assertions matching evidence |
| ATP Budget Awareness | Plans an action (e.g. complex reasoning) | ATP balance query | If ATP < action cost, defer or simplify | Sustainable resource usage |
| Trust Calibration | Trusts peer at 0.8 (high confidence) | Coherence index of peer behavior | If coherence drops, reduce trust proportionally | Adaptive trust relationships |
| Confabulation Detection | Generates a claim about past behavior | Search context window for supporting evidence | If no evidence found, flag as uncertain | Honest reporting with calibrated confidence |
The Universal Pattern
Every feedback loop follows the same structure: believe something, check it against reality, decide based on the check, and act accordingly. This pattern applies to thermostats, immune systems, scientific method, and AI meta-cognition alike. The insight from the SAGE research is that AI systems can learn to close these loops autonomously — but only above a certain capacity threshold. Below it, we must build the loops into the architecture.