Files
engine/enso/Cargo.toml
zeekay 4611304b3a fix(enso): refuse a featurizer whose width is not the policy's
Featurizer::dim() was declared and never called. policy::D is a compile-time
constant, so a wider featurizer was accepted in silence and bilinear simply read
the first D values: x = [1; 16] and x = [1; 16] ++ [9999; 240] score identically,
240 dimensions discarded without an error. A learned feature head wired in that
way would not fail -- it would route confidently and wrongly.

with_pieces now asserts feat.dim() == policy::D, which makes the dead method
load-bearing at construction and leaves the hot path a bare dot product. new()
delegates to it so there is one construction path rather than two.

tests/seam.rs pins it, and fails against the unmodified code.
2026-07-16 19:27:22 -07:00

15 lines
429 B
TOML

[package]
name = "enso"
version = "0.1.1"
edition = "2021"
description = "enso: the learned, eval-trained, per-user-adaptive routing policy (the brain) for hanzo-router (the mechanism)."
repository = "https://github.com/hanzoai/engine"
license = "MIT OR Apache-2.0"
[dependencies]
hanzo-router = { version = "0.1", path = "../hanzo-router" }
serde = { version = "1", features = ["derive"] }
serde_json = "1"
[dev-dependencies]