One ecosystem. Explicit responsibilities.

The design takes a goal and its constraints through a reviewable IntentSpec to a Contract IR boundary, before solving, checking or refusing a plan. It keeps computational intent distinct from an agent’s economic commitment. The contract architecture and examples are documented; the parser, analyzers and Flow integration remain implementation work.

The language and compiler connect typed computation, canonical IR, supported verification procedures and target-specific execution. Deterministic output claims belong to a defined workload, arithmetic contract and tested target set. Native independence and broader target coverage remain active compiler work.

The governance layer evaluates supplied evidence, invariants and policy at the admission boundary. Parties or their agents agree terms upstream; witness recording and downstream settlement have separate responsibilities. Governance modules exist, with compiler and native execution integration still in progress.

Architecture scanning combines fixed-point metrics, rules and evidence-chain tooling to make structural change reviewable. Parser coverage includes Python, TypeScript, Go and Rust, with a separate targeted MIND reflection path. Support for all programming languages through explicit adapters, and pure-MIND self-hosting, remain roadmap work.

Flow develops the typed execution-graph layer that the MIND-Intent design is intended to produce. Graph definitions and composition are distinct from a running workflow; native execution and the CLI-to-runtime bridge remain integration work.

Agent components describe identity, capabilities, permissions and governed commitments. Computational contracts and economic commitments stay distinct, so a model’s proposed action does not grant itself authority. Runtime integration remains in development.

Persistent memory, retrieval and a propose–review–apply workflow carry context across sessions. Provenance, contradiction handling and retrieval receipts support inspection of what was remembered or changed, within the documented configuration and evaluation scope.

Intent classification narrows a large skill and tool library to relevant candidates. This preselection keeps working context bounded while NAESTRO coordinates the continuing trajectory.

Shared evidence interfaces connect computation, decisions and retained records. An evidence record describes what its producer observed; independent verification and operational trust require an explicit verifier and deployment contract. Adapter coverage varies.

A research prototype investigates field-based computation, persistent memory and long-context model architectures in MIND. Its empirical benchmark ladder remains open. Within the intent-to-action architecture, a model supplies proposals; the establishment and authorization decisions remain outside the model.

Research workflows organize proposals, experiments, failed attempts, retained results and review. They support investigation around the ecosystem with explicit evidence and acceptance boundaries.