Internal Technical Notes¶

docs/AI/ is the current implementation reference for maintainers and coding agents. User workflows belong in the Zensical pages at the root of docs/.

Active notes describe current contracts. The history/ directory contains dated migration and removal records; use those for rationale, not as the source of truth for present APIs.

Algorithms¶

• IPPO
• IQL
• Shared RL infrastructure
• Nash-Q
• Discrete PR2 trainers
• PR2-IQL joint-action scaling
• Handcrafted baselines
• Handcrafted rewards by environment

Environments And Objectives¶

• Native substrate architecture
• Native observation contract
• Objective and atom conventions
• Stag Hunt
• Coins
• Chemistry
• Harvest
• Territory
• Gift Refinements

Monitoring¶

• Counterfactual replay
• LTLf rewards and discounting
• Episodic reporting

Experiments And Verification¶

• Artifact persistence and partial resume
• Hyperparameter presets
• Timing sidecars
• Final satisfaction
• NashConv reporting
• NashConv auto-resume
• Verification rollout performance
• AlphaRank cross-play

Notebooks And Visualisation¶

• CLI overrides
• Algorithm toggle coverage
• Marimo compatibility
• Play-notebook globals
• Pursuit IPPO helper
• Pursuit multi-algorithm notebooks
• Timings report notebook
• Visualisation package
• Legend exports

Operations And Testing¶

• Cluster launchers
• HPC container toolchain
• Zensical site
• Curated API reference
• Test suites and runtime
• Pytest collection policy
• Matrix smoke-run tool

The one-time Pursuit export copier is maintained as scripts/migrate_pursuit_export_paths.sh; executable utilities do not belong inside the documentation tree.

Historical Records¶

• Early NashConv notes
• IPPO/IQL cross-repository audit
• Early Nash-Q research
• RL audit fixes, 2026-04-29
• StarCraft removal

Research Questions¶

• Open questions

Optimisations¶

• Reward optimisation change log
• Reward optimisation assumptions
• Reward screen: 2026-08-10

Maintenance Rule¶

When implementation changes, update the narrow active note that owns the contract and the corresponding user page when the behavior is user-visible. Keep paths relative, mark dated investigations as history, and run uv run zensical build --strict after documentation edits.