Mini Universe — Frequency-Band Navigation Q-Policy

A small, honest artifact from the Mini Universe (mini 宇宙) simulator — a custom multi-agent "expanding lattice" world, not a transformer/LLM. The only learned parameters are a single aggregated Q-table of shape (480, 6) — 480 grid states × 6 actions, shared across all agents. This is the "model" that is published here.

What it is

• World: each agent occupies a multi-layer lattice address and chooses among 6 actions (wait / frequency-up / frequency-down / expand-sub / expand-subs / refine) to navigate a logarithmic frequency axis split into 20 bands (floor 1e-19 Hz, ceiling 1e81 Hz, 5 decades per band).
• Learning: tabular Q-learning with UCB exploration and annealing, trained to reach a target frequency band and then dwell there.
• Weights: model.safetensors contains q_table (480, 6) — the policy averaged across agents (weighted by visit counts). This repository's checkpoint was trained for the goal band 10 (see config.json).

Files

• model.safetensors / pytorch_model.bin — the aggregated Q-table.
• config.json — hyperparameters (bands, state/action dims, trained goal & ticks).
• pu12_v33.py — the universe simulator (main loop, v3.3).
• pu12_v33_mirror.py — the training/evaluation driver (run, Lattice, Learners) and the benchmark probes.
• inference.py — load the Q-table and evaluate (greedy, no training) whether the policy reaches a target band.

Load & evaluate

pip install torch
python3 inference.py --goal-k 10 --device cpu
# -> {"goal_k":10,"reached":true,"ticks":...,"purpose_tick":...,"band_mean":...}

To retrain / reproduce, run python3 pu12_v33_mirror.py --B 1024 --train-ticks 80000 --goal-k 10 --device cpu (writes logs_mirror/benchmark.json).

Benchmark (honest)

A mirrored small-world benchmark scored 0.62 / 1.0: strong task learning and perfect internal energy conservation, but no positive transfer between goals (negative few-shot plasticity — catastrophic interference) and no language-level structure in action sequences. See the companion current-status.md / log.md in the source repo (github.com/logos-42/NODS) for the full write-up. Published as a reproducible baseline, not as a claim of general intelligence.