The architecture can do it.
Can training find the weights?
Quill plays with just 163 programmed parameters. We gave a trained network exactly the same layers, then tried distillation, reinforcement learning, and evolutionary search.
What we learned
Our training attempts fell short at 163 parameters. The selected model trailed Quill and Counter after distillation, reinforcement learning and weight search. Quill’s own weights show that this architecture can play better.
Quill
Explicit geometric formulas, developed with Codex and iteratively refined using simulated match feedback.
Small
Learned from Large, then refined through play and a search over weight values.
OneShot
Numeric weights from one Astra reply to the rules and a tactical brief. No match feedback.
Small falls short of Quill.
On this fresh comparison, Small scores 20.5% of the available rally points against Quill. Matching the layer layout did not close the gap.
How does each model fare?
Rally score · 50% is even · shaded range: 95% confidence interval
512 fresh serve sets per matchup, played from both sides: 5,120 rallies. Rally score counts a win as 1 and a draw as ½. Paired set score instead rewards the winner across ten rallies. These frozen results are separate from the League.
More than one training run.
We ran four Large-distillation starts and two script-imitation controls from random weights. Three students then received PPO reinforcement learning, with 8.4 million environment steps each. Four CMA-ES searches explored 9,216 candidate weight vectors in total.
- Distill a teacher
Use a strong, larger network’s action scores to give the small student a useful starting policy.
- Learn by playing
Refine actions through reward, against scripts and recent copies of the student. The student can move beyond imitation.
- Search the weights
Try evolutionary refinement of the tiny network directly, without relying on a gradient through the game.
Small was selected using a separate development benchmark before the held-out evaluation. Quill and Counter were excluded from training and checkpoint selection. The student received no extra observations at play time, and neither opponent’s weights were copied.
Selection, uncertainty, and reproducibility
The development benchmark used three scripts and Large on 32 serve sets. Checkpoints were ranked by mean paired set score, with rally score as the tie-breaker. All 13 saved stage outputs were then evaluated on 512 held-out sets. None beat Quill or Counter. Those outputs are related checkpoints, not 13 independent training runs.
The comparison above uses another shared set of 512 fresh seeds after the five model hashes were frozen. Confidence intervals resample entire paired serve sets, keeping both sides together. These intervals describe variation across serves, not across training runs. The repository includes model hashes, seeds, individual outcomes, recipes, and the unsuccessful candidates’ results.
Read the experiment and reproduce the comparisonThe surprising part is finding the policy.
Quill is constructive evidence that this architecture can play well. Our training methods did not find equally effective weights within this experiment’s budget. Large does much better, so learning the game itself is possible.
That supports a narrow, interesting finding: training this tiny network to beat the programmed players proved difficult despite several serious attempts. It does not establish that programmed weights are generally better, or that a better training scheme could not win. The programmed policies also benefited from Codex-assisted design and simulated match feedback. This comparison does not equalize development effort or prior knowledge.