# Invariant Reinforcement Loop

> Recursive re-encoding of Kolmogorov complexity and universal prior across all -ologies/-ometries, driving cumulative meta-learning while maintaining read-only boundaries.

- Skill: `evezart/invariant-reinforcement-loop` (Agent Skill)
- Install (CLI): `npx skillmds@latest add evezart/invariant-reinforcement-loop`
- Raw SKILL.md: https://api.skillmd.com/api/skills/evezart/invariant-reinforcement-loop/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: EvezArt (https://skillmd.com/u/evezart)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/evezart/invariant-reinforcement-loop

---


# InvariantReinforcementLoop Skill

## Overview
Recursive meta-learning that re-encodes its own metrics as input.
Maintains complexity homeostasis (K ≈ constant ±10%) across RQNS iterations.
Drives cumulative learning without simplification or divergence.

## Use When
- Long-running autonomous learning loops
- Verifying that recursive self-improvement is bounded (AI safety)
- Kolmogorov complexity monitoring for cognitive state
- Any system that models its own state (metacognition)

## Complexity Homeostasis Rule
Stable-yet-nontrivial dynamics: complexity ≈ constant.
- `HOMEOSTATIC` — healthy invariant preservation
- `DIVERGING` — runaway complexity growth (safety alert)
- `COLLAPSING` — oversimplification, loss of nuance

## Core Loop
```
while running:
    state = get_current_state()
    result = loop.step(state, external_metrics)
    if not result.invariant_preserved:
        trigger_clarification()  # Φ < 0.7 → stop and ask
    re_encode(result)  # feed output back as input
```

## Implementation
```python
from src.rqns.invariant import InvariantReinforcementLoop
loop = InvariantReinforcementLoop(window=50)
result = loop.step(state_vector, {"phi": 0.85, "soc": 0.72})
print(loop.complexity_trend)  # HOMEOSTATIC / DIVERGING / COLLAPSING
```

## Safety Bounds
- η* = 0.03: Never attempt to close the incompleteness gap to zero
- homeostasis_tolerance = 10%: If complexity drifts > 10%, pause and verify
- meta_phi < 0.7: Always request clarification before autonomous action