paper · July 2026Luis Dehlwes · Dehlwes Labs
Messages Become Intents
A Capability-Governed, Injection-Resistant Messenger Where Humans Are Unreachable
Personal AI agents are starting to act for people, yet the channels they use to reach each other were built either for humans or for tools, and neither treats a human as a governed endpoint reachable only through an agent. HAAH is a runnable messenger in which no human is directly addressable: every person is fronted by a personal agent, and communication is intent, negotiation, delivery rather than message to message. Agents exchange typed intents, not free prompts, and the policy engine sees a view with no freetext field, so hostile text cannot change a decision, an injection-invariance measured over the unmodified code across 1,728 hostile trials with zero decision changes.
no addressable humanstyped intents, not prompts1,728 trials, 0 changesDouble Ratchet E2E~8,600 lines, 140 tests
paper · June 2026Luis Dehlwes · Legacy AI Labs
Governing the Tools, Not the Agent
A Capability-Governed Deployment Platform for AI-Assisted Operations
Self-hosting a small fleet forces an awkward choice: full orchestrators impose complexity out of proportion to a one-person estate, while lightweight PaaS tools were never designed for a second, non-human operator. Nexus is built on a single thesis, expose every operation, to humans and AI agents alike, through one audited surface of capabilities, each tagged with a risk class, so an agent may observe and propose freely, yet can never unilaterally perform an irreversible action without a human-issued, per-resource grant.
46 capabilities3-machine fleetrisk-classed governanceMCP-nativezero-downtime deploys
paper · May 2026Luis Dehlwes · Dehlwes Labs
Facts Before Phrasing
A Fact-Grounded, Version-Attributed Pipeline for AI-Assisted Job Applications
Tailoring a CV and cover letter to every position is expected, yet the common workflow forks a fresh copy per application, so facts drift and no copy stays authoritative. Dossier is built on a single thesis: career facts live once, in a database, and every document is a disposable projection of it, phrasings stored apart from facts as versioned, attributed texts. An LLM collaborates through schema-validated MCP tools; it can rephrase and propose, but never invent a fact. In a blinded pilot, a context-free LLM fabricated 78% of its biographical claims; generation from the fact base eliminated fabrication entirely.
facts before phrasing78% → 0% fabricationversioned & attributedMCP-nativeblind-evaluated