Pong evidence directory

Latest: teacher-study-v1 preserves one completed teacher round, 12 replay-verified episodes and the failed second-round teacher call. Neither candidate was promoted; final tests were not run.

Future reviewed Pong runs belong under experiments/pong/<study-id>/. Keep small JSON indexes readable in Git; archives and videos use Git LFS. Working data belongs in ignored artifacts/pong/<study-id>/<round-id>/ until reviewed. These are storage conventions, not automatic CLI routing.

The human-readable Pong teacher log links proposals to evidence and decisions. Pong's evaluation profile defines game-specific meanings; each run also needs a frozen protocol.

Historical evidence stays at its original paths

• Question history: value/direct pilot programs and gates.
• Original pilot archive manifest.
• No-FIRE proposal, results and manifest.
• Fixed-frame controls and manifest.
• Local calibration and manifest.
• Long trial and manifest.

This directory adds a game-level entry point; it does not copy, rename or reinterpret those archives. See the existing publishing/replay guide for review, checksums and restoration. Verify remote LFS retrieval before claiming a new archive is preserved. Preserve failed work and clearly mark missing historical transcripts.

Question diagnostics v1 adds 16 replayed Python control episodes, original-state adherence analysis and 16 videos, with zero new model calls.