Docs

Documentation Index

ZeroProofML v0.6.0 hardens the SCM stack: bottom results now have stable fault, semantic, or mixed provenance; non-finite payloads fail closed; simplification modes are explicit; and projective gauge conventions are named.

The site opens on Getting Started. Use this index when you want the whole map.

Read First

  1. Getting Started - install, first SCM values, strict inference, and promoted examples.
  2. SCM Foundations - bottom semantics, payload-plus-mask tensors, projective tuples, weak sign, and fracterm flattening.
  3. Training Guide - projective heads, losses, gradient policies, coverage control, semantic targets, and dtype guidance.
  4. Inference & Deployment - strict masks, stable fault/semantic provenance, tau_infer, schema-v2 ONNX bundles, fallbacks, and monitoring.

Reference

  • API Reference - stable versus experimental import surfaces.
  • Performance & Benchmarks - scientific benchmarks, microbenchmarks, reports, and run artifacts.
  • Development Guide - extending APIs, debugging masks, testing, logging, and example labels.
  • Experiments & Reproducibility - paper replay, benchmark harness, reference robotics deployment, and archived workflows.
  • Integrations - NumPy, PyTorch, JAX, ONNX Runtime, C++, REST, ROS 2, and visualization/reporting adapters.
  • Release Notes - readable v0.6.0 change summary and compatibility notes.

Library Docs Mapping

The library MkDocs site uses many numbered raw pages. This curated site keeps the same topics, but folds them into fewer task-oriented pages:

Raw library-docs area Curated page
Getting started, SCM foundations, projective learning Getting Started, SCM Foundations
Gradient policies, loss functions, adaptive coverage, debug logging Training Guide
Strict inference, tau_infer, masks, stable provenance, bundles Inference & Deployment
API stability, namespace policy, utility support boundaries API Reference
Benchmarks, optimization, artifact glossary Performance & Benchmarks
Verification notes, FRU placement, example status Development Guide
Reproduce-the-paper, experiments, reference robotics, composability Experiments & Reproducibility
NumPy/Torch/JAX, ONNX Runtime, C++, REST/gRPC, Triton, ROS 2, visualization Integrations
Changelog Release Notes

Stable Contracts To Remember

  • Use zeroproofml.* for new code and docs.
  • Keep bottom_mask authoritative; do not infer bottom status from payload values.
  • Stable strict inference unpacks as (decoded, bottom_mask, gap_mask).
  • Eager strict-inference results also expose stable fault_mask, semantic_bottom_mask, and bottom_provenance attributes with bottom_mask == fault_mask | semantic_bottom_mask.
  • Schema-v2 ONNX bundles ship the same six outputs; TorchScript keeps the three-field tuple.
  • Non-finite P, Q, or validity-factor payloads route to fault_mask (fail-closed).
  • Strict fracterm/FRU simplification is the default; simplification_mode="field_rational" is an explicit unsafe opt-in.
  • numerical_hazard_threshold is monitor-only metadata; finite tiny denominators no longer set fault_mask.
  • For paper reproduction, pin zeroproofml==0.4.3; v0.6.x is the active development line.
  • Benchmark claims should come from versioned run artifacts, not hand-curated notebook summaries.