モデルが提案する。実行が統治.される。証拠がエクスポートされる。

Mindburn Labsは自律システムのProof-First実行インフラを構築。

PIPELINE ACTIVE0 evaluated

Request flows through governance → cryptographic receipt issued

LIVE
success: OPEN SOURCEHELM OSS v0.1.1
npx @mindburn/helm
Verifies the latest HELM OSS release offline. Prints SHA-256 + receipt.
accent: CONFORMANCE0 checks
L1Structured
24/24
L2Receipted + Chained
17/18
L3Verified
10/12
conformance targets

Governed Execution Pipeline

Every AI action is proposed, policy-checked, executed, and cryptographically receipted. Hover any stage to see its invariant and proof output.

InvariantAction is structured as a typed proposal — never raw execution
EmitsCallProposal { schema, params, caller_id }
Kernel boundary
Proof surface
External

Research Workstreams

Six workstreams drive verifiable AI execution forward. Each produces standards, code, and cryptographic evidence.

A2AAgent-to-Agent ProtocolsEVDEvidence & CryptographyKRNKernel OS ArchitecturesVERVerification Tooling
HELM Kernel
Core execution kernel, proposal-centric architecture, proof graph construction.
Evidence Systems
Evidence packs, hash-linked receipt chains, verification tooling.
Conformance
L0–L3 conformance levels, test suites, CI enforcement gates.
A2A Protocols
Agent-to-agent communication standards, interoperability, transport.
TITAN Organism
Reference system for governed trading — TypeScript brain, Rust execution.
Verification Tooling
CLI verifiers, visual regression, security scanning, demo contracts.

リリースとリサーチ

2026年2月success: Release
HELM OSS v0.1 — パブリックリリース
完全なオープンソース実行カーネル。OpenAI互換プロキシ、ProofGraph、fail-closed PEP。
▸⊢ View Evidence / Verify
hashab494e896e
verifynpx @mindburn/helm --id helm
2026年2月accent: Research
ProofGraph仕様 (RFC-001)
改ざん防止、ギャップのない実行トレイルのためのPrevHashチェーンとLamportClock単調性。
▸⊢ View Evidence / Verify
hash851587be3d
verifynpx @mindburn/helm --id proofgraph
2026年2月warning: Engineering
Bounded Compute — WASI/Wazero (RFC-004)
ネットワークなし、ファイルシステムなし、計算制限付きのサンドボックス実行。
▸⊢ View Evidence / Verify
hash0ea5f24987
verifynpx @mindburn/helm --id bounded

References & Proofs

1. Fail-closed by default ✓ Verified

2. Zero-dependency setup ✓ Verified

3. Offline-verifiable receipts ✓ Verified

4. 5–10 ms overhead per tool call ✓ Verified