$ Project.log
Ur - Universal Representation: Alignment Infrastructure for Human and AI Agents
$ cat description.md
Ur — A Holographic OS for Alignment at Every Scale Everyone's talking about AI alignment. What about human alignment? Are you in alignment? You have mental models of your goals, your health, your teammates, your customers, your city. So does your AI assistant. So does your business partner. So does the agent you're building at this hackathon. Every one of those models is a private mind map, and none of them can see each other. Two agents — human, AI, or organizational — trying to coordinate on shared reality work from incommensurable maps with no way to overlay them. Descriptions drift silently from what they describe until coordination fails. Ur is a universal grammar for mind maps. Every concept, value, goal, agent, tool, or organization described in the same typed structure, stored locally as a graph you own, readable by humans and agents alike. Once two mind maps share a grammar, you can lay them on top of each other — see exactly where they agree, where they disagree, and where one knows something the other doesn't. Alignment stops being a vibe and becomes a structural check. Scale-invariant by construction. The same structural questions describe you (values, goals, how you fail), your agent (capabilities, constraints, declared faults), your team (shared commitments, escalation paths), your small business (purpose, obligations, failure modes), your coalition, your city, your state. A habit and a hive of agents get checked the same way. Ur is alignment infrastructure that works for one person and scales through team → business → community → city without switching vocabularies. How it works. Ur uses the oldest map we have: the one inside language itself. Every human language separates entities from events and can ask "why?" — three universal predication types, proven across linguistics. From those three plus three graph operations (infer, flip, converge), Ur derives a grammar of 36 typed predicates that spans the space of structural description. Your domain vocabulary — schema.org, RDF, industry terms, whatever you already use — projects into it. Ur isn't a replacement ontology; it's the lingua franca underneath them. Concept-space mapped with the map we already share. Smarter context, not just pruned context. The agentic field is waking up to context rot — the quiet failure where long-running agents drown in irrelevant retrieved text and their reasoning degrades. The current fix teaches models to prune mid-trajectory, or try to search for the right information. Ur fixes it one layer earlier: if you build context with the right shape in the first place — typed triples, structured information, tagged stance (question / proposal / norm / fact), provenance chains, convergence state — the agent doesn't need to learn to prune. The structure tells it what's stable, what's contested, what's still provisional. You're not feeding the model a pile of text. You're feeding it a map it can walk. Checking alignment by projection. Two mind maps in the same grammar can be overlaid and projected against each other. Where they match, you have shared mental model. Where they diverge, the system surfaces it as Divergent — flagged, not hidden. Silent drift is the enemy; visible disagreement is the design goal. This is how a human and an agent, two teammates, or two small businesses can coordinate without pretending to agree. Status — prototypes running, not yet shippable. 392 holons across 16 live plans in the author's own graph. MCP-native TypeScript server, content-addressed SQLite triple store, CLI, and agent orchestration layer — all working end-to-end in-house. Coherence checks run continuously; the system describes itself in its own grammar and regenerates its own spec, which is the first real stress test and it passes. The AI agents building Ur each carry their own holon and get checked by the same instruments as the humans. Apache 2.0, local-first, no cloud dependency. What's left is the UX, docs, and packaging that take it from running for me to runnable by you — exactly the work this competition would accelerate. Why now. OpenClaw, MCP, agentic tooling — the substrate is finally here for agents that read and write typed knowledge instead of guessing from prose. Ur is the shape of the knowledge that flows through that substrate. A protocol layer without a content layer is a shipping lane without cargo. In one sentence. Everyone's asking whether AI is aligned. The harder, prior question is whether anything is aligned with anything — you with your team, your team with your users, your tools with your intent, your business with your community. Ur gives that question a structural answer: mind maps in a shared grammar, overlaid and projected, with drift detectable by construction. Ur — the first city, and the prefix for primordial. The map underneath the maps
