Pong research

Pong is the current test environment for the teacher-driven question optimizer. Its small action space makes it a useful starting point, but neither mastery nor repeatable multi-round learning has been established.

See the first teacher study results: one completed round, both candidates rejected, and a technical stop before a second proposal. The next research design separates strategy quality from Jev question execution. No new live study is launched by that document.

Start here

• Evaluation profile: Pong outcomes, diagnostics, stopping rules and the distinction between policy and value-prediction studies.
• Teacher log: each proposal's evidence, change and disposition.
• Approved longitudinal teacher study: three rounds, a no-trajectory-feedback control and final evaluation; see its frozen execution protocol and diagnostic-parser continuation.
• Published evidence: locations of replay data, original model exchanges and archives.
• Shared research: optimizer design, evaluation framework and roadmap.

Directory contract

docs/pong/
  README.md
  evaluation.md
  teacher-log.md
  teacher-rounds/<entry-id>.md
artifacts/pong/<study-id>/<round-id>/    # Local, ignored working evidence
experiments/pong/<study-id>/            # Future reviewed JSON and LFS archives

The directories describe different views of the same experiment. A teacher entry links to exact questions, context, evidence and results; it does not replace raw records. New studies should use these game-specific paths. Existing reports, videos and archives retain their original paths and checksums. Their links below provide one Pong entry point without rewriting historical evidence.

The CLI already accepts explicit output paths, but does not automatically route commands by game or create the new human-readable teacher entries. Until that integration exists, create an entry with the shared template before a new proposal and publish its outcome after evaluation.

Current experimental reference

The latest long trial used pong-vertical-control-v2, Jev jev-1.13.0, all six native actions and probability argmax. It reached 7:18 at the 20,000-frame cap on development seed 56. That run is unfinished. V2 is a selected experimental reference, not a policy promoted by the earlier five-point pilot gate; that gate rejected promotion because its candidate windows were incomplete. V3 is not promoted.

The automated study runner now records fixed-input comparisons and isolated teacher proposals. Its first study retained v2 in both arms and stopped on a failed teacher invocation in round two. Final tests were not run. See the linked result report.

Historical studies

Study	Report	Teacher change?
Value-question pilot	Offline prediction and selection	Yes; rejected for MAE regression
Question representation	Score/Choice comparison	Representation experiment; not a consecutive teacher round
Direct-policy pilot	Vertical-control proposal	Yes; not promoted by its gate
Fixed-frame controls	Four policies, four seeds	No; existing programs
No-FIRE follow-up	Manual wording revision	Yes; development-informed, not promoted
Local match calibration	Nine local episodes	No Jev or teacher calls
Long Jev trial	20,000-frame feasibility	No; unchanged v2

Observation and replay contracts remain in observation.md, protocol.md and replay.md. Historical run protocols stay frozen; this directory does not alter their selection gates or results.

The question execution diagnostics protocol freezes the next offline adherence analysis and separate literal Python strategy trials. No new model calls.

Question diagnostics results: low interception-rule adherence, 16 separate literal-control episodes and no new model calls.