Roadmap
Roadmap
What we are working on next. Dates are targets, not commitments — every item ships behind a design doc and a benchmark.
Q1 2026
- Shipped
Rust reader crate (0.6)
Zero-copy .nsz reader for services that aggregate benchmark results outside Python.
Rust SDK → - Shipped
Cloud-storage streaming
Range reads against S3, GCS, and Azure Blob for .nsz artifacts larger than local disk.
Cookbook →
Q2 2026
- In progress
Activation attribution probes
Gradient-based attribution as a first-class probe kind. Output is compatible with existing comparison metrics.
- In progress
Interactive playground
In-browser Pyodide + wasm reader for exploring artifacts without a local install.
Playground →
Q3 2026
- Planned
Distributed recording
Coordinated recording across data-parallel ranks with a single output artifact.
- Planned
SavedModel adapter
First-class support for TensorFlow SavedModel exports, matching the JAX and PyTorch adapters.
Q4 2026
- Planned
Report-as-artifact v2
Signed comparison reports with an embedded provenance chain suitable for regulated environments.