Research provenance and implementation decisions

September 18, 2026 ยท View on GitHub

Historical research snapshot: private memorysaver/idea commit d591120290ef6216c1518e2c0d20bc0e5d8c7272, topic deep-research/jev-structured-question-rl/atari/ (2026-09-17).

On 2026-09-18, the owner designated this repository's docs/ as the canonical home for all project research, including the teacher-optimizer objective, evaluation methods, optimization patterns and paper planning. The research index, direction, optimizer design, evaluation framework, roadmap and related work consolidate that direction in English. Readers do not need access to the historical private notes. Those notes remain provenance, not a required upstream source of new research.

This repository owns the prototype code, tests, research and implementation decisions. It is not a copy of the private workspace configuration. No upstream configuration, credentials, runtime inventories, or ROM assets are included.

Implemented choices

  • Use Gymnasium/ALE directly and a small attributed Pong extractor instead of installing the whole OCAtari stack. Its inspected requirements pinned older NumPy/OpenCV versions and pulled in unrelated ML dependencies. The selected implementation is locked with Python 3.12, Gymnasium 1.3.0 and ale-py 0.11.2.
  • Implement both RAM and simple RGB object modes, but use RAM for the first controlled evaluation. Keep the mode in every artifact.
  • Start with fixed heuristic-continuation Monte Carlo branch labels and offline question selection. TD/Q-learning is deliberately a later, distinct experiment.
  • Support imported teacher proposals so the first experiment only requires a Jev key. OpenRouter is an optional teacher adapter, not a prerequisite for local baselines.

First-party references inspected

Official API claims and launch speed benchmarks are not evidence of this prototype's learning efficiency. Local environment and mock validation are recorded separately in validation.md.