Initial project portfolio

Data that remembers. Research that keeps its receipts.

The two projects share an evidence discipline, not a runtime dependency or marketing bundle. Each has a distinct scope, license, threat boundary, and research record.

Public / Apache-2.0

Hyphae

One process owns one local data directory and exposes bounded relational SQL, native structures, lexical search, and vector search over a shared transaction, WAL, MVCC, recovery, and proof substrate.

Primary formOne Rust binary

SurfacesEmbedded, local, HTTP

LicenseApache-2.0

View source Capability matrix

Boundary. Hyphae does not include replication, clustering, built-in TLS, at-rest encryption, a hosted control plane, an embedding model, or an LLM. “Proof” means verification against the documented commitments and rules; it does not establish that input data is true.

Public / Alpha / MIT

Hyphae Transformer

A from-scratch PyTorch framework for controlled experiments in identity-initialized residual learning for decoder-only transformers. Core provides reusable gates, model components, training utilities, and signal diagnostics; Lab provides typed hypotheses, staged experiments, explicit budgets, evidence-backed memory, immutable run manifests, and reproducible reports.

FrameworkPython 3.12 + PyTorch

Research axisResidual control

LicenseMIT

View source Research program

Boundary. Hyphae Transformer is research software, not a production model service or a claim that one residual strategy is universally superior. Current evidence is tied to the published architectures, corpora, seeds, token budgets, hardware, protocols, and stop decisions.

Software 03 / Inference

Hyphae BitNet

A correctness-first CPU inference runtime for supported ternary networks. It keeps packed GGUF as the durable store, creates an optional ISA-specific compute image under a fail-closed RAM cap, and separates prefill GEMM from decode GEMV.

Current product surfaces include one-shot generation, a native HTTP server, an experimental C ABI, JSONL benchmarks, exact I2_S and Q1 accumulation contracts, and an optional process-separated Hyphae gateway.

Explore Hyphae BitNet View source

Shared thesis

The shared interface is durable evidence.

Hyphae makes data, transactions, recovery, and proofs inspectable. Hyphae Transformer makes hypotheses, experiment budgets, manifests, checkpoints, metrics, failures, and reports inspectable. Both treat operational history as part of the artifact rather than disposable exhaust.

The projects remain independently usable. Hyphae contains no language model, and Hyphae Transformer does not require Hyphae. Any future integration must be proposed and evaluated as an explicit technical contract.