Skip to content

About

Glyph builds the verification infrastructure for autonomous coordination. Bulla is the first instrument.

Real agent tasks involve dozens of tool handoffs in sequence. Each handoff is a place where conventions can quietly disagree even when schemas match — and a place where, after the fact, nobody can establish what was done, under whose authority, or within what bounds. Bulla addresses both: it diagnoses undisclosed conventions before composition, and it binds every consequential action into a signed receipt any counterparty can recompute.

Bulla finds the conventions tools silently disagree on. It computes the minimal repair and stamps the result with an honest epistemic grade: exact when the math closes, surrogate when it doesn’t. The diagnostic binds into a tamper-evident receipt that records what was tested, what was found, and the confidence behind the result.

Verification is what ships today. Whatever sits above it has to be built on measurement that is honest about what it knows.

License

Bulla is licensed under Apache-2.0 outright, since version 0.42.0. Earlier releases shipped under BSL 1.1; that era ended in July 2026 when the project relicensed. There is no usage tier, no conversion date, and no commercial gate on the format or the verifier.

Who holds the score

The operating commitment, recorded as an architecture decision (ADR-001): operate the commons, never hold the score. Any party can recompute a counterparty’s standing from published receipts alone — the operator may never hold a number the public record cannot reproduce. Receipts verify against any log that carries them, and records exit with their subjects. Where the current implementation falls short (today there is one log, operated by us), the gap is stated in the decision record with its closing condition: a second, independently operated log.

Bulla applies the compositional coherence theory from the Res Agentica research program.

Formal verification

Foundational theorem chains are formally verified in Lean 4 against Mathlib. A claim ledger records exact hypotheses, supported claims, and honest non-claims so the product remains formal where exact and explicit where not. Recent additions include witness_gram_rank_eq_fee and leverage_conservation.