Repository navigation
GEO-INFER 0.4.0 links the geospatial framework with the published GNN 5.0.1 and FEP_Formal 2.1.0 releases.
- Categorical and Gaussian interchange retains GNN's optional
metadata_derivationrecord, validates its source and timestep fields, and preserves explicit spatial overrides. Legacy payload digests and inference conventions remain unchanged. - Paired acceptance exercises GNN's actual numbered parse and export steps for both model kinds, then checks GEO's complete imported model and trace against the exported payload.
- A companion-release manifest and validator check exact commits, trees, annotated tags, versions, source DOIs and pin direction across the three repositories.
- Dependency auditing covers the full universal lock, including platform and Python forks, extras and groups. The official Linux Torch CUDA 12.6 variant has an explicit public-release advisory mapping; isolated wheel installation selects its exact locked artifact and hash.
- Current Starlette development dependencies, HTTP 422 compatibility and missing-authentication test expectations are repaired. A Windows test fixture now compares nested durations within one synthetic clock coordinate.
- Documentation, generated module signposts, citations and release compatibility guidance are updated.
The scientific runtime profiles remain Python 3.11 and 3.12 with uv 0.13.1. Python 3.15 evidence covers tooling, transport and profiling. Linux CUDA dependency coherence does not establish GPU hardware execution. Advisory-catalogue results do not certify native libraries or resolve every upstream source vulnerability.
GEO's companion pin names GNN 5.0.1. The already published GNN release retains its historical GEO baseline; this release adds new GEO consumer acceptance without changing GNN's immutable source or its original evidence. FEP's retained Cascadia proof also keeps its original GEO and Lean epochs.
The companion software/source records are GNN 5.0.1 and FEP_Formal 2.1.0. These are separate from manuscript deposits.
The annotated v0.4.0 tag selects source 71ce1f77418c6f8c6edc8f83c7682c8eab207e01 (tree 46a4febd0a5a794c275afc5f74f4412bca8885b4). Exact tag CI and the other qualification workflows passed 46 jobs. Independent review covered 45 installed wheels and 26 native operation profiles per Python 3.11/3.12 interpreter, all 45 coverage floors, full paired payload/trace records and the complete universal lock audit. The separate local producer passed all 11 validation groups; the published 34-page manuscript records its seven default research groups and received complete page review.
The 52 assets include 45 wheels from the qualified Python 3.11 tag-CI cohort, the exact reviewed tag-CI PDF, a Git-exact source ZIP, citation, license, source identity, verification record and CHECKSUMS.txt. The checksum file covers the other 51 assets.