Portable R and JVM inference

CTBoost ships source packages for scoring prepared numeric JSON-predictor artifacts in R and Java 17. They are inference runtimes, not training bindings, and they do not alter the model's conditional-inference feature selection or split points.

Supported profile

Both runtimes implement the repository's json-predictor-prepared profile for format versions 1 and 2. They validate the envelope, quantization arrays, finite model values, tree dimensions, child and feature indices, reachability, and acyclic traversal before scoring. Duplicate JSON keys are rejected.

The input must already have exactly num_features numeric columns in the order used by the exported model. Artifacts with expects_prepared_features=false are rejected before the scorer reads fitted-pipeline fields. Use the Python JSON runtime when inference must reproduce CTBoost's fitted categorical, text, or embedding preprocessing.

JVM

Build and test the Java 17 module:

cd bindings/jvm
mvn test

Load JsonPredictor from the package io.github.ctboost.inference, then call its raw, probability, or class-prediction methods. The module is source-distributed and has not yet been published to Maven Central.

R

Build and check the pure-R package:

R CMD build bindings/R/ctboost
R CMD check --no-manual --no-build-vignettes ctboost_*.tar.gz

Use ctboost_load_predictor(path) followed by ctboost_predict(model, data, type=...). The package is source-distributed and has not yet been published to CRAN.

Security and conformance

The JSON model must be trusted. Structural validation and traversal limits make malformed artifacts fail closed, but the model fingerprint is a consistency checksum, not a digital signature. The shared fixtures under tests/export_conformance bind Python, R, and JVM raw/probability/class outputs for regression, binary, and multiclass models.

Training APIs, JNI/native bindings, raw fitted preprocessing, and model-registry authentication are outside this first portable-inference slice.