Release Notes

This page condenses the v0.5.1 changelog into the changes most relevant to users of the public docs. For implementation-level history, keep the repository changelog as the source of truth.

v0.5.1 Highlights

ZeroProofML v0.5.1 extends the SCM-first stack with stronger deployment, reporting, benchmark, and integration surfaces while keeping the stable strict inference contract unchanged:

decoded, bottom_mask, gap_mask = result

Documentation

v0.5.1 adds a broader published-docs set:

reproduce-the-paper guidance
examples inventory and promoted tutorials
focused task guides for tau_infer, masks/provenance, deployment bundles, ROS 2, and visualization/reporting
clearer namespace guidance for zeroproofml.* versus compatibility zeroproof.* imports
documented stable versus experimental API boundaries

This curated docs site folds those raw pages into fewer task-oriented guides.

Deployment And Bundles

Important deployment additions:

load_onnx_runtime_bundle(...) and ONNXRuntimeBundle
run_bundle_reference_smoke_test(...)
richer metadata.json schema fields for tensor signatures, batch-axis semantics, mask semantics, preprocessing/postprocessing IDs, and normalization sidecars
strict_inference_exports metadata for distinguishing current merged-mask bundles from future provenance-bearing contracts
C++ ONNX Runtime example and header wrapper for stable strict bundles
validation report summary sidecars

ONNX remains the preferred deployment path. TorchScript support is legacy compatibility only.

Monitoring And Provenance

v0.5.1 adds more operational visibility without changing stable outputs:

StrictInferenceMonitor.export_state(...)
structured strict-inference event logging
optional bottom/gap/fallback trigger records
opt-in experimental provenance diagnostics
fault-versus-semantic bottom-rate reporting when diagnostics are supplied

The Q2 provenance decision keeps richer provenance as experimental. Stable consumers should continue to depend on decoded, bottom_mask, and gap_mask.

Training, FRU, And Examples

Training and modeling additions include:

documented loss curricula and logging conventions
experimental FRU expression helpers for local P/Q flattening
FRU equivalence tests and strict-check demo
clearer guidance that flattening is a post-training analysis/export validation tool, not a per-step training mechanism
maintained quickstart, coverage-control, bundle-export, and FRU tutorial examples

Benchmarks And Reproducibility

The benchmark and paper-replay path is more auditable:

paper replay bundle under artifacts/paper_2026/
make reproduce-* shortcuts
benchmark resume, skip-complete-seeds, and force-rerun controls
versioned run manifests, provenance, per-seed result schemas, and artifact hashes
validate_run_dir(...), load_benchmark_run(...), and compare_benchmark_runs(...)
regenerated benchmark, bundle, and training-log reports
DOSE operating-point, Pareto, and diagnostics artifacts
RF frequency-response traces and qualitative figure packs
IK/robotics report figure packs when saved diagnostics are present

Integrations

v0.5.1 adds or documents:

ROS 2 companion workspace bootstrap
strict-inference ROS node and lifecycle variant
ROS 2 StrictInferenceResult.msg
QoS presets for low-latency control and offline replay
telemetry vectors and CSV exporters
RViz marker overlays for workspace and RR IK debugging
REST adapter decision: useful as a thin optional wrapper over validated bundles
Triton-style serving decision: downstream recipe candidate, not a first-party runtime path yet

Compatibility Notes

zeroproofml.* is the canonical public namespace.
zeroproof.* remains a supported compatibility namespace for existing code.
Experimental outputs, plotting helpers, provenance diagnostics, and downstream simulators may change faster than stable SCM, training, inference, and benchmark APIs.
Old benchmark artifacts without current schema markers should fail fast instead of being silently mixed with current claim runs.

Docs