GNN 4.1.0

GNN 4.1.0 aligns public execution contracts, strengthens owned filesystem and process boundaries, and improves scientific presentation, diagnostics and installed-package acceptance. The 25-step workflow and supported scientific semantics are preserved.

Changes

• Python, CLI, REST and MCP execution share admitted steps, frozen selections, typed backend options, run identity and current artifact receipts. Invalid executing-step requests refuse before output creation; an empty selected model view remains valid no-work. Installed APIs use explicitly configured model/output storage independently from installed code. Read the migration guide.
• Logging, website generation and RxInfer interchange/analysis have coherent owners with preserved public consumers. Typed failure reasons retain source/path context. Logging handlers cannot mutate each other’s message; malformed native CSV rows cannot misalign valid traces. Round-trip testing uses the actual supplied source and distinct saved format artifacts, including concurrent same-name models.
• Saved Markdown types and nested boolean arrays retain their authored values, dimensions and descriptions through both readers and JSON/XML/YAML reopening. Existing literal size/depth guards and unsupported-value recovery remain intact. GUI 1 edits preserve canonical sections and newline styles; ontology labels take precedence over bare mappings.
• Generated Python preserves authored display strings, quoted names and JSON keywords as data. Canonical PyMDP and DisCoPy consumers retain their existing scientific contracts; separately exported generators remain separately qualified. Canonical PyMDP refuses nonfinite numeric metadata before creating or replacing generated files; the migration guide documents this finite-JSON requirement for both modes. Generated source compilation and metadata checks do not establish inference readiness.
• Native JAX command-line streams, generated result files and saved execution logs use UTF-8, preserving authored Unicode across source, paths and outputs. Scientific algorithms, supervised process boundaries and original deadlines remain intact. Windows cancellation targets the actual direct worker and refuses unsupported descendant containment.
• Saved JSON/YAML models retain explicit time-step sizes and canonical datatype values. Installed YAML parsing accepts leading comments and reports malformed syntax through the original failed parse result. Registry/convert refusal and the schema facade's visible Markdown recovery remain separate contracts. Formal-format saved consumers preserve authored values and provenance; this does not certify external compilers or proofs.
• Gradio builders support the declared installed dependency and actual callbacks preserve canonical transition axes, signed/zero vectors and float values through save/reopen. Read the GUI migration guide before regenerating older rounded, malformed or transposed exports.
• Saved current-schema ActiveInference.jl outputs recover their own CSV observations, actions and beliefs without borrowing another model’s implementation path. Malformed numeric rows are rejected as complete rows, preserving alignment and the available signed EFE artifacts. These are saved-file consumer checks, not new Julia execution or scientific equivalence claims.
• Processing reports refuse missing or failed required outputs. Pipeline aggregation honors requested formats, retains actual partial files and preserves historical identities when replaying saved runs. MCP stdio retains fragmented/coalesced frames and drains admitted work at EOF; read failures remain failures in saved round-trip counts.
• POSIX creation, leases and deletion use descriptor-relative operations with explicit trusted-parent limits. Cancellation verifies owned cleanup and stream drainage; unverified cleanup retains records and artifacts. Windows direct-worker cleanup is declared precisely and stronger descendant guarantees refuse before launch. These are the documented process/filesystem boundaries, not a hostile-code sandbox.
• Scientific figures identify controls, axes, iteration/timestep bases, probability/entropy quantities and reported uncertainty. Unknown units and conventions remain explicit. Native categorical and Gaussian traces, selected animation frames, actual stored artifacts and source hashes support finite visual review.
  Canonical action planes use previous-state/next-state axes. Small correlation views use a symmetric domain derived from actual finite data; all-zero display scaling preserves the original numbers.
• The full 552-production-image encoder comparison found 5.48% less median encoder CPU time with identical decoded pixels, dimensions and metadata, at a 3.07% file-size increase. This is encoder-stage evidence on the measured native environment, not a whole-pipeline or cross-platform speedup. Repeated, concurrent and bounded localhost distributed witnesses retain actual timings, resource observations and admission limits.
• The .agent_rules directory is modularized into 22 guides with task-based routing and dedicated workflow, run ownership, scientific evidence, interface, CI, documentation, security and release modules. Stable paths remain available; current registries, configuration and executable gates own changing facts. Source archives exclude generated import-linter caches; archive verification binds package and documentation bytes to Git. Maintained guidance uses canonical package paths and live owners. Additional behavior tests exercise authored formats, public entrypoints, native artifacts and failure paths. Separate comprehensive coverage traces Python subprocesses while preserving existing core reports, scopes and gates.

Verified delivery

Annotated tag v4.1.0 identifies accepted main 9255e19f951d. All 25 required source jobs in ten workflows passed on main and again at the exact tag; the actual source-bound dependency graph also passed. Independent raw coverage exceeds 80% separately on Python 3.11, 3.12 and 3.13, with 8,859 distinct selected nodes per environment. All five installed native platforms passed 163 environment-case executions with zero failures, errors or selected skips. Exact CI JUnit selections, ordinary wheel/sdist Git and RECORD parity, all 22 agent guides, outside-checkout installed imports and manuscript producer custody were verified. The macOS tag job alone was retried after a hosted-runner capacity failure; the four original passing lanes and actual producer-attempt custody are retained.

FEP main 34f8f3330dbf passed its 20 required jobs, native 168-topic Lean catalogue, declaration audit and 388-page render acceptance. GEO main c1027a9d538f passed all 23 required checks, including both native interchange matrices pinned to this GNN source. Current issues, Dependabot and code-scanning alerts are zero. Full evidence and artifact hashes are in release-verification.json and SHA256SUMS. All 13 attached release assets were downloaded from their public URLs and verified against the accepted bytes, provider digests and checksum files. The post-publication receipt records the actual publication, original public downloads and current gates.

The version-specific Zenodo DOI 10.5281/zenodo.23245091 resolves to the published 4.1.0 source archive. Its provider checksum, all 3,289 tagged Git blobs and the released manuscript were verified directly. The source-archive receipt includes exact source and manuscript hashes and ZIP metadata qualifications. This automatic source archive does not include the separately attached wheel, sdist or checksum list; full distribution archival remains E5 and requires archive-owner connected authorized access.

Remaining limits

Released THRML 0.1.4 still produces NaNs for the independent inactive-padding/structural-zero witness on both supported JAX splits; strict model rejection remains. Upstream issue closure is not a compatible released fix. The directly verified 4.1.0 version-specific source archive matches this release, while archival of the separately attached distribution assets remains an external requirement. The concept DOI does not establish distribution-asset parity. PyPI publication and major scientific/LLM extensions remain separate tasks.

Python 3.14 GUI callbacks remain blocked by an upstream Gradio deprecation under the unchanged strict warning policy. The full Python 3.11, 3.12 and 3.13 GUI lanes and Python 3.14 builders do not certify callbacks on 3.14. Theme warnings remain visible. Complete native reports independently exceed 80% on all three supported Python environments; the comprehensive guard now enforces that threshold with the original selectors, denominator and strict warning policy.