Atari challenge scope

Our goal is to attempt and work toward mastering every discrete single-agent ALE/*-v5 game registered by the locked ale-py installation. On 2026-09-18, ale-py 0.11.2 exposes 104 such games. The ROM package and website may list a different total; neither defines this experiment's denominator. Modes, difficulties, continuous actions and multiplayer are separate extensions.

jev-atari games generates the inventory from registration. games --check boots each game and executes legal actions without model calls. The checked inventory is environment-check-2026-09-18.json. All 104 passed eight-frame smoke checks. This only tests startup and stepping.

Define success separately for each game before declaring mastery: not every game has a final level or a universal completion event. Record mode, difficulty, score or completion criterion, evaluation budget and replication. Report coverage of attempts separately from games meeting those criteria.

The second project goal is to discover reusable teacher-optimizer patterns. A game used to develop an adapter or optimization method is not automatically a held-out transfer test. See the research roadmap for that distinction.

Capability	Pong	Other registered games
Boot/action smoke check	Passed	Passed in recorded installation
Generic raw-RAM runner	Available	Available
Semantic object adapter	RAM and simple RGB rules	Not implemented
Live Jev experiment	Recorded pilot	Not yet tested
Value-based labels	240-frame first-point outcome	Not implemented
Multi-round learning or solved-game claim	Not established	Not established

Adding a game properly

Create a game directory such as docs/pong/, using the shared evaluation-profile template and teacher-round template. Keep each game's endpoints, diagnostics and mastery criteria explicit instead of inheriting Pong's metrics.

1. Record environment, ROM hash, legal actions, mode and difficulty.
2. Define observable state, missing values, history and startup/life-loss behavior.
3. Validate action effects; joystick labels alone are not semantic evidence.
4. Run local controls, including a simple domain rule where available.
5. Freeze a direct question and bounded Jev experiment before interpreting outcomes.
6. Define a separate value target; do not reuse Pong's labels blindly.
7. Publish inputs, outputs, teacher changes, all attempts and a replay report.

The raw-RAM runner is an exploration route, not object understanding. It uses ALE default mode/difficulty, zero startup no-ops, no automatic FIRE, and a raw-frame cap. The agent must start/resume games through legal actions. Compare matching protocols; scores across games are not directly comparable.

Source: ALE environment reference.