munish0838/duck-engineer-rounds

Beam as an RL engineer on the Microduck Arena (sprint-2m)

Every round of an experiment asking: can Beam, following Pollen Robotics' own agent guide, engineer a training recipe that beats the reference sprinter at the same training budget — and does it predict its own results correctly?

Each row is one round: the exact context Beam was shown, its raw reply and reasoning, the diff applied to the recipe, the validation outcome, every checkpoint's score from Pollen's official Arena evaluator, and the token cost.

19 rounds across these arms: A, A2, C, C2, C3, C4, C5, C6, C7, S1, S2.

The best round so far scored 2.745 s (5/5) — S1 round 8.

How one round worked

1. The context pack was built: Pollen's sprint_2m/AGENTS.md verbatim, the pristine env.py and challenge.toml, the reward API extracted from the installed packages, the budget, the bar, and the session's history.
2. Beam replied with one JSON object — analysis, hypothesis, knob values, exact-once search/replace edits, an optional new rewards.py, and a prediction of its own score.
3. The reply was applied unedited to a fresh branch off the pristine recipe, then checked: schema, paths, exact-once edits, knob ranges, a banned-pattern guard, Pollen's own pytest suite, and a CPU smoke test. A failure went back to Beam verbatim, as a retry within the same round. It was never fixed by hand.
4. A human operator read and ran the printed training command: from scratch, 4096 envs, 1000 iterations, seed 1, one L4 GPU.
5. Checkpoints 250/500/750/1000 were each exported with publish --dry-run and scored with arena score. A run's result is its best checkpoint: 5/5 finishes first, then the lowest time.

Every round reset to the pristine recipe, so a change Beam wanted to keep had to be proposed again — which makes each round's diff self-contained.

The bar is Pollen's published reference: 4.883 s, 5/5 finished, at 0.55 m/s (task Mjlab-Sprint2m-MicroDuck).

Fields

field	what it is
prompt_system, prompt_user	the exact context Beam was sent that round
prompt_est_tokens	its estimated size
analysis, hypothesis	Beam's own words; hypothesis became the commit message
change_type	knob, env, reward or combined
knobs, edits, new_files	the change, exactly as Beam specified it
expected, expected_score_s	Beam's prediction, before the run
reasoning, reasoning_tokens	Beam's thinking, when the provider returned it separately
diff	what git saw after the change was applied
validation_error	why a reply was refused, when one was
commit, branch, ckpt_repo	the code that trained, and where the checkpoints went
checkpoint_scores	per checkpoint: the 21-speed sweep, the 5 seeds, the score
score_s, seeds_finished, command_vx, best_checkpoint	the round's result
prediction_error_s	score_s − expected_score_s; positive = Beam overclaimed
report	the report that went back to Beam next round
warnings	auto-detected trouble: a penalty term that pays, reward farming, falls
tokens_in, tokens_out, beam_wall_s, job_minutes	cost

Also in this repo

• rounds/<arm>/r<k>/ — every file each round produced, including each Beam attempt's raw request and response under attempts/.
• OPERATOR_LOG.md — every command run, with its outcome and a timestamp.
• context/ — the static context pack Beam was shown, and SOURCES.json recording the exact package versions and commit it came from.
• PROJECT_BRIEF.md, BEAM_CONTEXT_SPEC.md — the experiment's design, written before it ran.

Honest limitations

Simulation only. Few runs. GPU reinforcement learning is noisy from run to run, which is why the experiment includes a control on the same budget and a second-seed confirmation — any difference smaller than that seed spread is noise, not a result. One event, one model, one provider. Local Arena scores are indicative; only the Arena Space times an official entry, and simulations on different CPU architectures drift apart after a few seconds.

Credits

Pollen Robotics for the Microduck, the challenges repo, AGENTS.md, the Arena, its evaluator and the reference policy. Reflection for Beam. The helper scripts were built with Claude Code; every recipe change was proposed by Beam, and a human operator ran every step and approved every job. No hand-tuning.