When \(\mathcal{R} \ge 1\), there is surplus; when it falls below 1, the pattern eats itself. Truth, boundary, and trust are how agents keep the balance.
What is it like to be an information-persisting system that is learning to understand itself, its environment, and how to survive in noise, danger, and competition?
This project answers that question in three tiers: a formal law (\(\mathcal{R} \ge 1\)), a book on phenomenology, and software that instantiates the same fractal hierarchy from tokens to a nation-scale trust ledger.
| Tier | Question | Where |
|---|---|---|
| Theory | What must anything do to keep existing? | papers/information_persisting_systems.md |
| Experience | What is it like to be such a system? | books/book1/ |
| Implementation | How do you run the loop in silicon? | Aion LLM → Aion Core → Aion Blockchain |
| Level | Fractal node | Repository |
|---|---|---|
| L1 | Neurons / tokens | aion-llm — FractalGPT, per-token \(\mathcal{R}\) |
| L2 | Neural network | aion-llm — FractalLM, usage.persistence |
| L3 | Collection of agents | aion-core — Loop, Processor, Machine, Market |
| L4 | Trust ledger (nation / currency) | aion-blockchain — Proof of Trust |
Operator map: COMPONENTS.md · integration walkthrough: round-trip
Any bounded pattern that stays distinct over time must keep its persistence ratio \(\mathcal{R} \ge 1\): predictive income must pay for noise, model error (\(\mathcal{D}_{KL}\)), and fatigue (\(\Gamma\)), with shelter \(\Psi\) from above and substrate integrity \(\Phi\) from below.
Not a separate substance — the running state of a node that has learned to keep \(\mathcal{R} \ge 1\) from the inside. The Useful Approximations Framework (UAF) in books/book1/ gives it functional shape:
aion-mini is stage 0 boot — not a fifth layer. It joins the PoT chain, brings up aion-core, and activates the cognitive processor before the Loop takes over.
FractalLM is the compression layer: information formed upstream in aion-core and aion-blockchain is distilled into candidate checkpoints first, with LoRA planned next and full retraining after that. Chat completions expose usage.persistence back to the higher layers for scoring.
Sense–plan–act–score: Loop, Processor, Machine, and prediction markets. This is the main task-level information-formation layer, where contact with reality becomes structured traces, beliefs, norms, and learnings. It also does the first compression pass at runtime — notes, state mode, process structure, and a recommendation engine that learns which tools, processes, and files belong in each step’s context before those traces are later moved into weights.
A git-native chain whose consensus is paid in KL bits: voters cast probability distributions, those beliefs are resolved against outcomes, and the validated society history flows back into core and eventually into model training.