Skip to content

Release Readiness

This page is the package-facing release checklist for SILVA Networks. It is designed to catch citation drift, documentation gaps, missing notebooks, and broken package builds before a public release.

Current Release State

Item Status
Package version 1.2.2
Citation metadata CITATION.cff, README, references page, and BibTeX file updated
Zenodo archive all-versions DOI: 10.5281/zenodo.21770098; v1.0.0 DOI retained in release history
PyPI package release target: silva-networks==1.2.2
Release workflow a validated v* tag builds distributions, creates the GitHub Release, and publishes through PyPI Trusted Publishing
Docs source widget top-right GitHub facts refresh against source_release_tag: v1.2.2 before cached release data is reused
Article citation arXiv:2607.28989, submitted July 31, 2026
Article PDF docs/assets/papers/silva-networks-arxiv-2607.28989.pdf
License MIT
Book and solutions manual Planned long-form learning assets
Implementation registry silva_networks.coverage.implementation_cases()
Cortex hierarchy vector and spatial SILVACortexLayer transitions, heterogeneous SILVACortexNetwork points, and SILVAImageCortexClassifier documented and validated
Point architecture catalog ten vector, token, and spatial modules with shape, gradient, fixed-point, tiny-data, example, and notebook checks
Generalized cases sequence, multiscale vision, Jacobian, IGNN, INR, diffusion, and coupled RAFT/DEQ-Flow APIs documented and validated
Advanced equilibrium families monotone graph, injected transformer, Poisson mirror, physics-informed ODE, implicit DAE, and adversarial residual mechanisms documented and validated
Full-scale family surface all 64 canonical families have data, literature, benchmark, scale-control, and extension guides; compact dense/scalable equivalence checks are executable
Learned solver and backward policies HyperDEQ, exact implicit gradients, JFB, SHINE, phantom gradients, and unrolling are documented with separate forward/backward controls and analytic checks
Circuit equilibrium family QDEQ provides a differentiable compact statevector route, external measured-circuit adapter, image contracts, direct/implicit schedules, and Jacobian diagnostics
Guided navigation all 210 navigable Markdown pages and 83 rendered notebooks include contextual next steps and an extension/reproduction path
Canonical notebook curriculum all 109 package, bridge, and unreleased book/research notebooks include executable extension, equivalence, gradient, reproduction, and scaling material
Portable test suite 385 core and extended tests pass with no skipped outcomes; device checks use CUDA when available and CPU otherwise
Public datasets UCI tabular loaders plus TorchVision adapters for MNIST, FashionMNIST, KMNIST, EMNIST, CIFAR10, CIFAR100, and SVHN
Public results Results records measured validation metrics, tensor shapes, residuals, and reproduction commands
CLI workflow CLI Guide, silva-experiment, silva-download-datasets, scripts/smoke_test.sh, config listing, config display, device override, and dotted --set overrides
Release audit script scripts/release_audit.py
Notebook validation script scripts/run_notebook_smoke.py

The release checks establish package behavior, numerical equivalence on compact problems, notebook execution, documentation integrity, and distribution installability. They do not claim that every cited paper benchmark has been rerun at its original compute scale. A benchmark result is publishable only after its official data, split, preprocessing, metric, and compute protocol are run and archived separately.

Release Checklist

Check Command Must pass
Release audit python scripts/release_audit.py yes
CLI validation bash scripts/smoke_test.sh yes
Python lint ruff check src tests examples scripts yes
Unit tests pytest -o addopts= -rs yes; every collected test must pass and none may be skipped
Extended tests pytest -o addopts= tests_extended -rs yes; every collected test must pass and none may be skipped
Strict docs build mkdocs build --strict yes
Public results page silva-experiment --config <config> --output-dir outputs yes for published validation rows
Notebook validation python scripts/run_notebook_smoke.py --all --timeout 180 yes; all 109 canonical notebooks must execute
Package build python -m build yes
Distribution metadata twine check dist/* yes
PyPI trusted publisher PyPI publisher for jseluis/silva-networks, workflow release.yml, environment pypi yes
Zenodo archiving Zenodo GitHub integration enabled for jseluis/silva-networks; v1.0.0, v1.1.0, v1.2.0, and v1.2.1 archived, and v1.2.2 metadata prepared yes
Release trigger push the validated annotated v1.2.2 tag after main CI and Pages pass yes

For local offline validation, python -m build --no-isolation is acceptable when the active virtualenv already contains the build requirements.

What The Release Audit Checks

The release audit is intentionally conservative. It fails on missing package surface, missing citation assets, outdated citation language, navigation gaps, and version disagreement. It warns, rather than fails, on optional local tooling such as command-line PDF utilities.

Area Check
Citation arXiv ID in README, CFF, references page, and BibTeX
Release metadata .zenodo.json, PyPI workflow, Zenodo DOI, PyPI URL, and publishing instructions present
Docs source widget mkdocs.yml source release tag matches the package version and the Material source-facts cache is guarded
BibTeX silva2026silvanetworksstructuredimplicit and software entries present
Docs required learning, API, experiment, and audit pages exist
Navigation new release/audit/derivation pages are wired into mkdocs.yml
Package pyproject.toml and src/silva_networks/__init__.py versions agree
Coverage each implementation case points to docs, notebooks, tests, and public objects
Test completion test sources contain no skip markers; optional integrations must test either the installed implementation or the explicit missing-extra contract
Companion assets book/manual page lists the planned long-form learning path
Stale text no outdated citation wording remains

Release Notes Template

Use this structure when cutting a public release:

SILVA Networks 1.2.2

Article:
Jose Luis Lima de Jesus Silva. SILVA Networks as Structured Implicit Layers and
Vector Attractors via Dynamic Interaction Fields. 2026. arXiv:2607.28989.

Package highlights:
- Structured SILVA equilibrium layers and presets.
- Cortex hierarchies with independently configured per-point architectures and a ten-entry vector, token, spatial, attention, and spectral catalog.
- Sixty-four canonical family routes with explicit data contracts, primary sources, benchmark paths, scale controls, and extension points.
- Fourier, graph-physics, homotopy, distributional, monotone graph, generative transformer, Poisson mirror, physics-informed ODE, and implicit DAE equilibria inside SILVA.
- Matrix-free derivatives and Newton-Krylov stages, fused or chunked attention, factorized graph maps, and chunked measure discrepancies.
- Lazy tensor shards, distributed loading, mixed precision, gradient accumulation, and complete checkpoint resume state.
- ODE, implicit PDE-step, Poisson, and Fourier-operator derivations connected directly to SILVA fields and solver diagnostics.
- Picard, Anderson, Broyden, and GMRES diagnostics.
- DEQ, MDEQ, Neural ODE, optimization, and optical-flow bridge material.
- Configurable sequence, multiscale, graph, INR, diffusion, and coupled RAFT/DEQ-Flow cases.
- Exact implicit, finite-unrolled, and phantom gradients with indexed trajectory supervision.
- Solver-learning, backward-approximation, quantum-equilibrium, source-equilibrium, evidence, and experiment-protocol APIs.
- Package-native notebooks, examples, public experiment cards, numbered references, citation audit, and full-scale training guidance.
- Companion book and solutions manual roadmap.

Validation:
- python scripts/release_audit.py
- python scripts/run_notebook_smoke.py notebooks/*.ipynb notebooks/package_api/*.ipynb notebooks/implicit_bridge/*.ipynb --timeout 180
- pytest
- mkdocs build --strict
- python -m build
- twine check dist/*

Known Warnings To Report

Warning Meaning
Article PDF metadata differs from arXiv Re-download https://arxiv.org/pdf/2607.28989 and rerun the release audit.
Companion book/manual PDFs Planned learning assets tracked by the public roadmap.
Optional PDF command-line tools missing Minimal machines may omit these tools; pdfinfo is enough for the package release audit.
Isolated package build cannot download build requirements Use python -m build --no-isolation in a prepared local virtualenv, or run the isolated build where package indexes are reachable.
Vision extras may be absent Install .[vision] only when running torchvision-specific examples.
Full cited benchmarks are not part of the CPU release audit Use the family guide to obtain the source protocol, then archive the resolved configuration, dataset version, hardware, checkpoint, and metrics for the selected study.
Vision dataset downloads are larger than tabular downloads CIFAR and the full TorchVision suite are opt-in public checks; cache them under data/ and keep generated data out of commits.
CUDA validation is hardware-dependent Device tests always run and select CUDA when available; release performance claims and CUDA-specific kernels still require a CUDA machine.
Full real TorchVision suite is opt-in The package tests the runner route and real CIFAR10 smokes locally; the complete real image suite should run where dataset archives can be cached.
Optimization extras may be absent Install .[optimization] only for CVXPYlayers experiments.

Where to Go Next

Question Page
How can I run every required check? Run Everything
What is the publication sequence? Publishing
Which documentation changes are recorded? Documentation Log