Gus — selected 757k depth-campaign specimen

757,184 parameters; 15 blocks, width 64, two heads. The registered population campaign finished at its final depth's operational plateau. Selection used the fixed 256-hand development panel, before the fresh confirmation panel was evaluated. The complete formation history remains archived.

Confirmation and scope

On the reserved 512-hand G0/Zeb mixed-table panel, this model's team won 45.70%, versus G0's 45.90%. The paired difference is -0.20 percentage points, 95% whole-hand bootstrap interval [-2.73, +2.34] (2,000 resamples, seed 42). This is one candidate-controlled seat, fixed mixed partners/opponents, bid 30, pip declarations and legal argmax. It is not a general tournament rating. The separate full-game comparisons were two-game integration smokes only.

The stopping rule was four generations without a two-percentage-point development gain, at depths 2, 4, 8, 12 and 15. Best development weights, including a depth's identity-grown initialization, remained eligible. Identity-grown initialization selected: True. An operational plateau does not prove a capacity ceiling. Neuron meanings remain for separate interpretation.

The selected specimen is an identity-grown initialization: its inherited learned policy is preserved, and the newly appended blocks have zero output projections. Its parameter count does not mean those added blocks learned useful features. The last fully trained final-depth candidate is also preserved, along with every intermediate formation checkpoint. It may have a lower development score than the selected specimen.

Inputs, outputs and learning

The original own seven tiles, public contract, relative seats and public play history are unchanged: 46 input tokens, 68 positions, 28 tile logits. There are no hidden hands, identities, planner values, legal-action lists or match scores in the input. Own legal moves constrain output selection. No bidding, Nello, value or splat head is present.

Each generation collected 4,096 hands and used policy-only PPO on actual team make/set returns. Other seats sampled 25% pinned Zeb and 75% uniform among the latest five completed Gus generations. All generations and formation checkpoints are retained, including the initial pretrained-v0.

Load and reproduce

Pin Hub revision campaign-selected-20261018. Use lem.observer.model.Observer(Config(**config)) with config.json and model.safetensors, or pass gus.pt to lem.observer.play.ObserverPlay. These are custom PyTorch classes, not a Transformers AutoModel. example.json contains an exact input and expected CPU float32 logits; tokenizer.json defines the game tokens. The adapter checkpoint supports torch.load(..., weights_only=True).

Experiment and protocol. Full formation and confirmation evidence. Weights and generated artifacts: CC0-1.0, Jason Yandell. The nanoGPT-derived implementation retains MIT attribution to Andrej Karpathy and Jason Yandell. The April oracle-era Gus is a different model lineage.