Module policy 

The policy shell’s ONNX contract, version 1: how a trained policy’s model must look so one generic program runs it in every arena.

• Inputs: one per declared observation, named exactly as declared, shape [1, *shape], f32. A u8 value is fed as f32 / 255. A policy may use a subset of the observations.
• Memory: an optional state_in: f32[1, S] → state_out: f32[1, S] pair, S at most MAX_STATE. Zeros at the start; each turn’s state_out is the next turn’s state_in.
• Continuous actions (declared f32): one output per declared action, named as declared, shape [1, *shape], in [-1, 1]. The shell maps each element with to_bounds: clamp to [-1, 1] (NaN is 0, the midpoint), then affinely onto the element’s declared bounds.
• Discrete actions (declared u8): one output per declared action, named as declared, of logits shaped [1, *shape, n], where every element shares one range [min, max] and n = max - min + 1. The shell takes the argmax of each element’s logits (argmax) and adds min.
• Metadata: META_CONTRACT, META_ARENA, META_PAYLOAD_SCHEMA, META_TRAIN_VERSION and META_OPSET, checked against the manifest.
• An input named NOISE is reserved.

botstadia-train implements the same map; the shared golden file (policy-golden.json, beside this crate’s tests) pins both.

Constants

CONTRACT_VERSION

This contract’s version, as the botstadia.contract metadata states it.

EXPORT_OPSET

The opset the export helper writes.

GOLDEN

The golden cases pinning to_bounds, feed_u8 and argmax, which botstadia-train checks its own implementation against.

MAX_STATE

The largest memory, in floats.

META_ARENA

META_CONTRACT

META_OPSET

META_PAYLOAD_SCHEMA

META_TRAIN_VERSION

NOISE

Reserved for sampled actions later.

STATE_IN

The memory pair’s names.

STATE_OUT

Functions

argmax

The index of the largest logit; NaN never wins, and ties go to the first.

feed_u8

An observation element as the policy sees it.

from_bounds

The inverse of to_bounds, for training: a declared value as the policy’s [-1, 1].

to_bounds

One element of a continuous action: a policy output in [-1, 1] mapped onto the declared [min, max]. Out-of-range outputs are clamped and NaN is the midpoint, so the result is always valid.