I'm Bartek Kus, a systems architect from Edmonton, Canada. I work on the trust problem in AI-native software delivery:
AI can write the code. The unsolved problem is trusting what it wrote.
My answer is architecting intent: the human authors the contract, agents do the work, and machinery (not optimism) refuses anything that drifts from the contract. Stop reviewing output; start constraining intent.
Everything lives under the statecrafting org; the public record is statecraft.ing, a static catalog of every specification in the family, pinned to immutable source revisions. The family is small tools with explicit boundaries, all Rust, all Apache-2.0. Each repository owns its own specs, and a library capability, a release and a deployed service are reported as separate claims.
spec-spine: a typed, hash-verifiable authority ledger over a markdown spec corpus. Every spec declares the files, sections and symbols it owns; a PR-time coupling gate refuses code that drifts from its owning spec. Byte-identical output across platforms, one exit-code contract, --json everywhere. Released: v0.32.0 on crates.io, npm and PyPI. It governs itself: its own coupling gate runs against its own spec corpus in CI.
statecraft-cli: a local environment for governed agent work. Register a repository, prepare an isolated workspace, supervise an agent through one unit of work, judge the result independently, keep a reviewable account. One machine, one repo, no account. Specified, implemented and tested; publication is deliberately deferred, so it runs from a checkout.
| Project | What it is | State |
|---|---|---|
| rahi | The Rust chassis for governed cells: co-deployed rauthy identity, in-process hiqlite Raft state, a hash-chained decision ledger, a deny-by-default capability kernel, an axum edge, single-container packaging with backup and restore. Apps compose it; they never fork it. | 0.7.1 release candidate on crates.io; supported topology N=1 |
| aicortex | Persistent, governed memory for AI clients over MCP, built as a rahi cell. Recalled content is always data; only a human-recorded promotion makes a memory instruction grade. | Specified; being built one spec per pull request |
| rustev | An embeddable engine for making, evaluating and recording decisions over partially trusted information: bounded plans, explicit abstention, an evidence record per decision. Proposals, never grants. | v0.5.0 on crates.io |
| wire-witness | Captures provider wire exchanges and produces attributed observations with custody facts. Testimony that never grants authority. | v0.1.0 tagged |
| action-gate · attest-ledger · canonical-keysort-json | The primitives: a deterministic Allow/Deny/Degrade gate over a check registry, a tamper-evident Ed25519-signed ledger with an independent verifier, and canonical JSON bytes at the hashing boundary. | 0.3.0 · 0.2.0 · 0.1.1 on crates.io |
- rauthy (@sebadob): a Rust OpenID Connect / OAuth2 / PAM identity provider with passkey-first security; the identity provider every rahi cell co-deploys.
- hiqlite (@sebadob): an embeddable SQLite that forms a Raft cluster for high availability and strong consistency; the replicated store inside every rahi cell.
- Architecting intent over vibe coding. No human reviews every line an agent produces; pretending otherwise just moves the bottleneck back to the human. Make intent the requirement, the requirement a spec, and the spec law.
- Agentic output is hostile by default. Agents earn passage by surviving gates, not by appealing to trust. A PR branch is a stranger's bytes until it is checked.
- Humans gate contracts, not diffs. Specs, approvals and irreversible boundaries are human territory; everything between them is enforced by machinery.
- A claim carries its grade. Specified, implemented, tested, released and deployed are different statements. A tool that cannot tell you which one it is making is not finished.
- Small tools, explicit boundaries. A gate does not log, a ledger does not decide, a witness does not authorize, a memory does not instruct. Each refuses the neighbouring job on purpose.
How I got here is one problem at three scales: a decade of digital identity (OIDC, SSI, DID) asking "can this person be trusted", platform engineering asking "can this system be trusted", and now governed agent delivery asking "can machine-generated change be trusted".
Rust • TypeScript • React • SQLite / hiqlite • PostgreSQL •
Kubernetes • Helm • Flux • Terraform • OpenTelemetry • Claude Code • Codex • MCP
➡️ https://bartekus.com ➡️ LinkedIn: /in/bartekus ➡️ Twitter: @bartekus ➡️ Email: [email protected]
"The human authors the law; the agents comply with it; the spine makes non-compliance impossible to merge."