Colony Kernel is deterministic mission control for AI-agent colonies. Every state change, budget, and human approval goes through an auditable, fail-closed core. An agent saying "this passed" is not a verification result, and nothing in the kernel treats it as one.
The problem it solves. Hand work to an AI agent and "I finished and it works" is a claim you cannot verify. A model that is wrong about its own success is worse than no agent at all. Colony makes the agent propose and a deterministic core decide. Trust comes from replayable evidence, not from the model's word.
From a 2026 study of 9,876 agent trajectories across 8 model families (arXiv 2606.09863). False success means an agent confidently reporting completion while the environment state says otherwise.
The thesis is not ours alone. Arbiter-K (arXiv 2604.18652, cure-lab/ArbiterOS) shipped a deterministic governance-first kernel first, and Rel(AI)Build (arXiv 2606.26924) converges on the same mechanisms. Colony's axis is mission control: budgets, approvals bound to transitions, durable rejections. Not security interception.
Full write-up →
The stage pairs below are projected from the kernel's own transition table, so you can press the illegal ones. Each button submits a forbidden move a compromised or overconfident agent might attempt, and the table answers with a reason instead of an exception. Every one of these is illegal by structure, not by likelihood.
One seeded demo mission, run twice during publication. Both runs are byte-identical, which is the whole point. Below is the complete, unedited output of the run that shipped this page (also served raw as demo.json):
loading demo.json …
Reproduce it yourself, offline, in seconds:
git clone https://github.com/faresrafat3/colony-kernel && cd colony-kernel
npm install
npm run verify # typecheck + lint + 147 tests + claims + demo twice, byte-identical
npm run demo > a.json; npm run demo > b.json
diff a.json b.json # empty — byte-identical replay
Default-deny edge table. Unlisted transitions are illegal, not "unlikely". Legality precedes storage.
Accepted events carry canonical hash + CAS version + monotonic sequence. Replay reconstructs identical state.
SHA-256 over RFC 8785 JCS bytes, bound to the exact transition, consumed ≤1, expired at commit.
Immutable limits with CAS rows. Agents can't mint resources or vote themselves a raise.
No self-mutation, no self-publication. Reserved public states are structurally unreachable.
Refused commands leave evidence and never mutate state. Audit trails don't have gaps.
Multi-agent frameworks let the LLM drive workflow state, so workflow is only as reliable as the model. Colony flips the trust boundary: models suggest, a deterministic kernel decides. Verification is proven by replayable evidence, not claimed by the model that wrote the code.
AI Agent Reliability Audit. A fixed-scope engagement for teams shipping AI agents whose outputs they can't fully trust. I audit your agent loop with the same doctrine Colony Kernel is built on: evidence over claims, replay over narratives, fail-closed over fail-open. You get a written findings report and a re-test tier.
Colony Kernel is the proof of work for this offer.
Book an audit (faresrafat3@gmail.com)