AI and ML tools and research, built by John Morrissey. Everything here falls into two piles: tools for making AI-assisted software engineering trustworthy, and research into reinforcement learning systems and the machinery that trains them.

A federation of code-intelligence tools for agentic software engineering, joined by a shared Stable Entity Identity (SEI) spine. Each member owns one domain — code structure, issue lifecycle, trust-boundary analysis, git/CI governance, change impact — and composes with the others enrich-only, never load-bearing. The product is the federation itself: the identity spine, the seams between members, and a hard provenance-honesty invariant (no result ships without its cause, reason class, and fix). The repo is the federation's documentation hub — doctrine, contracts, and the member roster live there.

A lifecycle-driven neural training system in which new computational structure is generated from the live state of a host network, compiled without semantic change, trialled in flash-cloned counterfactual branches, and admitted only if it beats a mandatory do-nothing alternative under a blind judge. Authority is split constitutionally across fourteen bounded domains, and admitted structure must keep earning its tenancy or be retired. Currently pre-implementation: the high-level design is complete and locked; engineering is beginning.

A world simulation engine for reinforcement learning and real-time applications, written in Rust with Gymnasium-compatible Python bindings. A tick-based core with pluggable spatial backends (1D lines and rings, square and hex grids, composable product spaces), a stateless propagator pipeline with write-conflict detection, ML-native observation extraction to flat tensors, deterministic replay with per-tick snapshot hashing, and arena-based allocation for zero-GC memory management. Published on crates.io and PyPI.

A pedagogical deep-RL environment designed to trick students into learning graduate-level reinforcement learning by making them think they're just playing The Sims. Agents survive by managing eight interconnected physiological and economic meters across fourteen affordances, with progressive complexity levels introducing partial observability, temporal mechanics, and intrinsic motivation. GPU-vectorized training in PyTorch, an adversarial curriculum from shaped to sparse rewards, and live visualization with episode recording and replay.

Deep reinforcement learning for Shogi. A Rust game engine and vectorized RL environment (exposed to Python via PyO3) paired with a PyTorch training harness using PPO, with a KataGo-inspired multi-head network architecture adapted for Shogi's drops, promotions, and action space.