Systems that remember. Records that hold.

Open research and infrastructure for local, durable, and accountable computation.

Hyphae Research Foundation is the institutional home being formed to steward Hyphae and related open systems research. We build software whose state, boundaries, provenance, and decisions can be examined rather than merely trusted.

The foundation is currently an open-source initiative in formation. It is not yet a separate legal or tax-exempt entity. Celiums Solutions LLC serves as organizer, interim custodian, funder, and operator until incorporation and documented asset transfers are complete, and as incorporator only if formation filings name it. It does not own the future foundation.

A local-first Rust data engine that unifies SQL, native structures, lexical/vector search, WAL, MVCC, recovery, and verifiable proofs in one binary. Public under Apache-2.0.

A PyTorch framework for identity-initialized residual learning research, with typed hypotheses, staged experiments, explicit budgets, immutable manifests, and reproducible reports. Public under MIT.

A correctness-first CPU inference runtime for supported ternary networks, with exact integer accumulation, model-family gates, bounded RAM compute layouts, and native CLI, C ABI, and HTTP surfaces. Public under Apache-2.0 with vendored and third-party portions retaining their identified licenses.

• Local first: useful without a required cloud or external authority.
• Evidence before headlines: performance and correctness claims carry protocols, source identities, limitations, and retained evidence.
• Failure is part of the contract: ambiguous input and exhausted budgets are refused rather than guessed.
• Negative results remain visible: failed approaches and corrected claims stay in the research record.
• Open means governable: code, decisions, roles, rights, and succession must all be legible.

We use conversation-gated contributions:

Discuss → Approve → Build → Review

Do not open a pull request before a maintainer approves a public tracking issue. Early ideas belong in Discussions. Every implementation requires an issue carrying decision:approved, a bounded scope, acceptance criteria, risk classification, and a committed reviewer.

AI-assisted contributions are welcome. Material AI use must be disclosed. The human contributor remains responsible for every line, claim, source, test, and artifact; project-specific review rules determine the required gate.

Read the organization-wide contribution policy before proposing work. Project repositories may add stricter technical, AI, security, and review requirements.

• Website
• Manifesto
• Research ledger
• Governance
• Participation
• Legal status
• Security

_{“Foundation” describes the institution being formed, not a completed legal or tax-exempt status. Project marks and third-party marks are governed separately from open-source licenses.}