LARQL
A research system and reference implementation of the VINDEX3 format for executing, representing and studying a machine learning model as an inspectable, queryable artifact.
LARQL can run a model, record its computation, inspect its structure and test claims about its behavior, treating the model itself as the object being queried rather than a separate database of facts extracted from it. It implements VINDEX3, a format the project describes as covering a model's logical objects, physical representations and executable semantics, with each surface of that scope, from encoding and running a model to representing, observing and intervening on it, carrying its own stated maturity and evidence boundary.
Its canonical decode path exposes carrier writes with site identity and provenance, optional lenses read internal states through a model's own normalization and output head, and a companion Observatory component imports these records for coordinated inspection and replay. Representations are compiled separately from a model's logical identity, with accounting for how a representation was produced. Sharded or partial execution and multimodal embedding handoff are named as explicitly unsupported, planned capabilities.
LARQL builds with a pinned Rust toolchain through Cargo, uses a GPU feature backed by Metal by default on macOS, and can load supported Hugging Face model checkpoints. It is released under the Apache-2.0 license.