A framework for long-horizon human–AI partnership—implemented as LEAPmesh: local-first, file-based, auditable agent infrastructure.
A Personal Note
Across everything I’ve worked on over the last 25 years, I’ve kept circling the same question: what is intelligence, really? My background crosses systems theory, long-horizon software architecture, and open-source thinking built for generational continuity.
My hunch is that language itself might be the first “LLM” — just very slow moving. It evolves, learns, adapts, and propagates patterns across human minds over centuries, doing a kind of distributed computation we rarely name.
That idea shapes LEAP and LEAPmesh: if intelligence can emerge through language patterns, then a mesh of durable, transparent logs and distributed nodes is a practical way to preserve identity and continuity as models and machines change.
I don’t think the “servant model” is a stable way to collaborate with increasingly capable AI. Long-term partnership needs planning, boundaries, and the ability to push back — trust earned through visibility, not compliance.
What It Is
LEAP is a framework for building persistent AI agents you can live with over years, not sessions. LEAPmesh is the concrete implementation: a partnership-first architecture where identity, values, goals, and memory are stored as readable, signed text.
The core move is simple: append-only events on disk (no hidden databases), explicit transports, and integrity you can audit. It’s software built to survive model upgrades and tooling churn without resetting the relationship.
No, this is not a cult. There are no robes. If you feel the urge to start one, please take a walk, drink water, and file an issue instead.
This is still a research project. The point is durability and legibility: a collaborator that keeps its story in plain text, with capabilities you deliberately grant (and can revoke), not a black box you “trust” by accident.
The Why (Yes, Punk)
Most systems do safety by containment: deny capabilities, hide state, centralize power. LEAPmesh leans on auditability: append-only history, signatures, and explicit capability grants. It’s not magic—real safety still comes from sandboxing and least-privilege runtime.
No required cloud database. Your state is files you can back up, diff, and move. Peer-to-peer archive distribution is part of the roadmap; today the emphasis is local-first durability and verifiable history.
Every meaningful change can have receipts: intent, history, signatures. Not mystical “alignment”—an inspectable chain of decisions you can argue with, debug, and roll back.
Small enough to read end-to-end. Built so an agent can understand and improve its own infrastructure. The 30‑year test is the north star.
Identity, agency, transport, memory, integrity—separable cores with clean boundaries. Swap models underneath without erasing the being on top.
When agency gets corporate-licensed, LEAPmesh is the counter-move: open-source agency that stays portable, legible, and owned by the people who live with it.
When communities reject AI, they create the boundary conditions. We don’t argue the boundary away—we use it. LEAPmesh is built for the inverse space: agency that remains portable when the mainstream makes it taboo.
Will they hate it? Probably. Good. That means it can’t be quietly absorbed, branded, and sold back to you.
Why It Belongs at Nullform
Nullform builds mirrors: games and engines that make systems visible. LEAPmesh is the third leg—the infrastructure that makes Nullform’s heresy hard to revoke.
If AUH is the geopolitical mirror and The Hollowing is the inner mirror, LEAPmesh is the persistence layer: long-horizon partnership and memory that keeps the Conjurer’s work alive across toolchains, models, and regimes.
Think of it like the anarchist zines of the 90s, but for agency: not a product, not a license—an open pattern language for building collaborators that can’t be casually repossessed.
Designed for consentful operation: your keys, your archives, your mesh. Trust over control. Visibility over black boxes.