Research direction

September 18, 2026 ยท View on GitHub

Updated 2026-09-18. This is the adopted project direction. Method proposals below are research questions, not completed learning results. Start at the research index for current evidence and document status.

Two goals

  1. The Atari challenge: work toward mastering every game in the pinned ALE scope with Jev, publishing reproducible attempts, limitations and per-game success criteria. A boot check, a short clip and mastery are different milestones.
  2. The optimizer research: discover how a teacher should turn environment experience into effective structured-question updates, and identify repeatable optimization patterns that generalize under a measured resource budget.

The second goal is a central research objective in its own right. A higher game score is an outcome to explain; the intended paper contribution is evidence about which optimizer mechanisms work, when they work, and what they cost. Negative results can narrow that explanation even when no game is solved.

The project's research ideas and paper planning live in docs/. The provenance record documents the transition from earlier private notes. Public research must be understandable without those notes.

System and learning target

A teacher proposes question changes from training evidence. Fixed-version Jev answers the questions; the environment supplies observed outcomes. A selector evaluates candidates and retains or rejects them. The learned artifact is the external question program. The initial study fixes both teacher and Jev weights.

Let theta be the question program and p_theta the returned action distribution. The current direct policy executes probability argmax, not a sample from p_theta:

policy_theta(observation, history) = argmax_a p_theta(a | observation, history).

Its objective is expected observed return under a predeclared evaluation protocol. Observations and history need not be a complete Markov state. Model confidence, rule agreement and textual explanations are diagnostic signals, not reward labels.

The teacher optimizer includes evidence selection, credit assignment, proposal generation, candidate selection, memory and budget allocation. Choosing a teacher model is one experimental factor, not the entire optimizer.

Two execution tracks

TrackQuestion outputLearning target and boundary
Direct policyOne Choice over available actionsImprove actual rollout return through question changes; current videos use this track
Value-basedStructured judgments about consequences of each actionLearn estimates tied to a specified return, horizon and continuation; requires prediction/decision validation

The original value-based idea remains central: strong reasoning turns experience into questions that a fixed executor can reuse, with a hoped-for cost advantage that still needs measurement. A structured critic may retain several judgments before aggregation, but their scales must correspond to an explicit target. Arbitrary confidence/cost/correctness scores do not become Q values merely by being added together.

The implemented critic predicts the first scoring event within 240 raw frames: the candidate action occupies the first four frames and a fixed Python heuristic controls the remainder. It is not a frame-240 position forecast or an optimal Q-function. Jev-continuation branching, learned aggregation, retrieval, and TD targets are separate future ablations. Neither text revision nor the current direct policy alone establishes Q-learning.

Falsifiable hypotheses

IDHypothesisRequired comparison or counterevidence
H1Experience improves teacher proposalsSame teacher and proposal/evaluation budget, with versus without trajectory feedback
H2An optimizer mechanism improves the searchSame teacher, executor, starting question and evidence access; ablate the mechanism and compare a relevant optimizer baseline
H3Credit assignment improves useful editsTrajectory-only feedback versus controlled action branches; charge all branch frames/calls and specify continuation
H4Structured representation helpsSingle guidance versus modular/structured variants with matched information and accounted execution cost
H5Improvements accumulate and transferIndependent optimization runs, regression cases, fresh seeds, then predefined held-out games
H6Jev offers a useful quality/cost tradeoffSame optimizer with another executor, plus direct Python execution where the rule can be implemented

H1 and H2 are the immediate research priorities. Study one factor at a time before combining mechanisms. The evaluation framework defines how to connect edits, behavior, outcomes and costs rather than treating any one of those as sufficient evidence.

Current evidence

The fixed-frame controls found better aggregate short-run return for the vertical Jev question, but only 66.85% agreement with its written 4px rule. The no-FIRE revision increased mean confidence and rule agreement while reducing aggregate return from -3 to -5. The long trial reached 20,000 frames at 7:18, with no native completion. These are useful control and measurement results.

We have not established a multi-round autonomous teacher optimizer, a repeatable experience-driven gain, full-match mastery, cross-game transfer, or a cost advantage. The interactive teacher has seen development results; a strictly train-only study requires a fresh, isolated teacher context and a recorded packet hash.

Route to a paper

The roadmap stages instrumentation, controlled optimizer studies, generalization and paper claims. Related work includes GEPA, TextGrad, Reflexion and TypeSafe's question-feature discovery example. Teacher-authored prompt revision is established prior art. A defensible contribution must be a measured mechanism, representation/credit result, generalization finding, or reproducible quality/cost result with appropriate baselines and uncertainty.