[package]
name = "op-fraud"
version.workspace = true
edition.workspace = true
rust-version.workspace = true
license.workspace = true
repository.workspace = true
description = "On-device fraud scoring. Pluggable Scorer trait; heuristic + ONNX + Burn implementations."

[dependencies]
op-core = { path = "../op-core" }
thiserror = { workspace = true }
serde = { workspace = true }
serde_json = { workspace = true }
time = { workspace = true }
# SHA-256 for hashing identifiers in features (no PII in model input).
sha2 = "0.11"

# ONNX Runtime binding for the production scorer. Pinned to 2.x.
# `load-dynamic` lets the operator place libonnxruntime.{so,dylib,dll} at
# a path of their choosing (App bundle on iOS, /lib on Linux kiosk, etc.).
# We do NOT enable `download-binaries`: operators audit and ship their own.
ort = { version = "2.0.0-rc.10", default-features = false, features = ["load-dynamic", "ndarray"], optional = true }
ndarray = { version = "0.16", optional = true }

# Burn 0.17.x is the pure-Rust ML alternative. NdArray is the only
# backend that compiles for no_std + WASM + iOS / Android, which is
# what the FFI layer (Phases 8-10) needs. We use Burn for operators who
# want to avoid shipping libonnxruntime altogether — kiosk-Linux builds,
# WASM deployments in browsers, or anywhere the binary-size budget is
# tight. ONNX Runtime adds ~10-30 MB; Burn's NdArray backend is ~1 MB.
burn = { version = "0.17", default-features = false, features = ["ndarray", "std"], optional = true }
burn-ndarray = { version = "0.17", default-features = false, features = ["std"], optional = true }

[dev-dependencies]
proptest = { workspace = true }
insta = { workspace = true }

[features]
default = ["heuristic"]
# Pure-Rust heuristic scorer. No external runtime.
heuristic = []
# ONNX Runtime scorer. Requires libonnxruntime at runtime.
onnx = ["dep:ort", "dep:ndarray"]
# Burn pure-Rust scorer. No external runtime, no extra shared library.
burn-backend = ["dep:burn", "dep:burn-ndarray"]