Approaches

The Ataraxos pipeline

How far does the paper's method get on one Radeon 8060S and this machine's 32 threads, and which of its parts carry the strength?

The mechanism

The training package implements the paper: a decoder-only setup network, an encoder-only move network with a key-query policy, self-play sampled directly from the networks with λ-returns and advantage filtering, the clipped loss with reverse KL to the data-collection policy and to the magnet policy, the power-law schedules for the regularisation coefficient and the learning rate, an EMA of the parameters, and a belief network trained afterwards on the final policy's self-play. Search applies one tabular magnetic-mirror-descent step over a thousand belief-sampled rollouts. The concept pages from self-play RL to test-time search describe each part with the paper's numbers.

The plan

1. A registered smoke run: twenty iterations of the small preset on 256 environments, evaluated against random and greedy. This tests the pipeline, not the method.
2. A scaled run at the largest network that keeps an iteration under a few minutes here, for as long as the entropy and clip-rate diagnostics keep the damping healthy, with a checkpoint evaluated against greedy and expectimax every ten iterations.
3. Belief training on the final policy, then the search comparison: network alone, network with search over the counting prior, network with search over the learned belief, each against expectimax on 200 paired games.
4. Ablations under the same budget: no advantage filtering, no setup learning, fixed regularisation instead of the schedules. Each is one registration.

Decision rule

The smoke run passes if the move network after twenty iterations beats random with a 95% interval above 50% effective win rate on 40 paired games. The scaled run is promoted above expectimax when it beats it on 200 paired games with the interval above zero at the same or lower time per move. Search is credited only if it beats the network alone under the same rule.

Evidence

Run on September 18, 2026: twenty iterations of the smoke preset trained in 43 seconds on the GPU without an entropy collapse (2.70 nats at the start, 3.08 at the end, never below the first value), but the move network scored 60% effective against random on 40 games with an interval of 44.6% to 73.7%, which covers 50%; 26 games stalled to the battleless limit, and the network scored 17.5% against greedy. The decision rule is not met; the rung stays unclaimed. See the report and the run record.

What could disprove it

The paper's damping schedules are tied to its iteration count and batch sizes; on a smaller run the same exponents may anneal too fast and collapse entropy. The diagnostic pair of entropy and clip rate is logged every iteration, and a collapsed run is recorded with its curves rather than restarted quietly.