Inspectable, CPU-first neural research.
Small dense, ternary and four-state decoders · exact synthetic curricula · reproducible training and evidence · measured geometry
Trinite is an open research workbench for asking what small, fully inspectable models actually learn. It keeps native numerical computation, admitted data, frozen protocols, checkpoint replay and PROVENANCE evidence separate from claims about reasoning or deployment.
Phases 0–4 are implemented. The source, test harnesses and bounded research experiments are usable; this is not a trained general-purpose assistant or a production-ready model. Phases 5–9 remain conditional on learning, replication, intervention and deployment evidence.
The first 27-cell dense/ternary/four-state comparison recorded 0% held-out complete-answer accuracy in every lane. Longer training passed some exposed-training criteria but failed important held-out tasks. Balanced arithmetic rehearsal restored training retention while maths performance traded off; an 8,192-update budget improved aggregate maths counts without any of the twelve maths tasks reaching its 90% gate. The later task-balanced experiment has a reported negative matrix in PR #29, but its complete original archive and independently retained commitments have not been located for verification. Archival closure remains unresolved.
The four-cell procedural-QEC v3 study also failed its training gate; complete replay and archival checks passed, while historical control ordering was not preserved.
Phase 4's initial item-versus-carrier geometry contrasts were negative. The E8/ETQ parity checks prove exact finite-payload agreement, and the activation-to-lattice experiment is retrospective and training-exposed—neither demonstrates learned geometric reasoning. Geometry loss remains zero. See the roadmap, Phase 4 limitations and initial milestone notes for the complete status.
Foundation, protocol and evidence checks use the Python standard library (3.11+). Model and training tests require the separately hash-locked CPU environment; no model weights are downloaded implicitly.
PYTHONPATH=src python -m unittest discover -s tests -v
# After installing the pinned CPU requirements:
PYTHONPATH=src python scripts/check_research_cpu.pyStart with the getting-started guide, documentation index and architecture. Implementing agents should read AGENTS.md, invariants and operating instructions. The original idea is preserved.
The immutable v0.1.0 tag is an initial research/source checkpoint covering 25 merged PRs numbered #1–#26 (excluding the unmerged revert PR #11), not a Phase 9 deployment certification, PyPI package version, released model-weights bundle or successful learning claim. The package metadata remains separately versioned until a deliberate packaging release. Keep immutable tags tied to the final reviewed commit; do not move historical evidence or tags.
Repository code: MPL-2.0. Admitted data, imported upstream references, external assets and future weights each retain their own rights and provenance requirements.