Question-only no-FIRE follow-up

This protocol is frozen before running the no-FIRE question. The exact proposal changes the guidance and name, retaining all six action options, their criteria, the observation format, 4px rule, and probability-argmax decoder. There is no action mask or confidence-triggered fallback.

Run Jev jev-1.13.0 on development seeds 36, 37, 46, 47, in that order, for 2,000 controlled raw frames / 500 decisions each, with no point cutoff. Use the existing RAM protocol: hold=4, sticky=0.25, reset NOOPs=0..30. Record all four videos, decisions, frame states, JSON exchanges and HTTP ledgers. One shared cap of 2,200 HTTP attempts includes retries for at most 2,000 decision calls. Errors stop execution and remain incomplete; native termination uses the existing zero-reward absorbing-tail convention. No teacher or extra diagnostic API calls.

The reference is the jev-vertical arm in the prior fixed-frame results, whose program hash is 2ef87250f491fcd5a2ba72f29e3f03b6bc66c4ab7315e9e724df76352326633d. That baseline is reused, not freshly sampled. These development seeds have already been inspected, and the proposal was authored in that context. This is an exploratory follow-up, not train-only learning, an independent final test, or a contemporaneous randomized comparison. API variability can affect differences.

Report every seed's scored/lost points and net-reward change, total cost, 4px rule agreement on visited states, NOOP/FIRE counts, mean confidence, fractions below 0.3/0.5, and the top-two probability margin. Confidence thresholds are descriptive, not validated action gates. The old run selected no FIRE variants in 4,000 decisions, so improved play cannot automatically be attributed to eliminating FIRE actions. No selection gate or automatic promotion is introduced.

uv run --env-file "$HOME/.config/typesafe/credentials.env" jev-atari compare-controls \
  --candidate-only --backend jev --model jev-1.13.0 --max-api-calls 2200 \
  --baseline-program examples/vertical-policy-program.json \
  --candidate-program examples/vertical-policy-no-fire-program.json \
  --seeds 36 37 46 47 --frames 2000 --video \
  --out artifacts/pong-no-fire-v3

--candidate-only executes only the candidate; it does not replay or invent baseline episodes. The inherited internal arm key is jev-vertical; the saved program name and hash identify this new v3 candidate. The required baseline-program argument is not evaluated in this mode. The runtime plan saves the actual source revision when --source-revision is provided. Use scripts/verify_controls.py to audit the four new episodes without API calls, then archive the original evidence using Git LFS.