QSOLKCB/Trinite：三元权重边缘 AI 研究项目进入 Phase 4 几何测量

An inspectable, ternary-weight edge AI research project: small models, curated data, geometric reasoning experiments, and PROVENANCE observation.

Status: Phase 4 geometry measurement implemented. Matched dense and ternary CPU decoders support native training, safe replay, detached prompt captures, and analysis with a pinned QSOL-GEO-REASON kernel and PROVENANCE observer. The three-lane/QEC comparison detour adds an opt-in classical four-state lane and exact generated QEC tasks before Phase 5. Its 27-cell run completed; all lanes scored zero held-out complete answers under the tiny frozen budget, leaving format selection inconclusive. The foundations increment adds independently checked arithmetic, folds and symbolic lattice tasks with a verified learning gate before Phase 5. Its 27-cell run verified, but dense training failed the gate, so held-out scoring stayed blocked. The training-only convergence sweep completed with no learning-rate candidate clearing all training gates. The subsequent longer training-only budget passed per-task training checks in all nine dense cells; held-out learning adequacy and Phase 5 remain gated. First visible-fixture geometry observations are descriptive; reasoning, intervention, and edge benchmarks remain pending. See Phase 4 evidence and limits.

Start with the documentation index, architecture, and roadmap. Contributors and implementing agents should read the operating instructions and invariants before changing the project.

Run the foundation and model checks with GETTING_STARTED.md. Foundation and evidence-verification commands use the standard library; model and training commands use the optional hash-locked CPU dependencies.

Repository license: MPL-2.0. Data and future model artifacts require their own explicit rights records.