XRAI ingests any stream — ARKit anchors, LiveKit packets, stock ticks, mocap, console logs, biosignals, LLM tool calls, news feeds, sim ticks — through a declarative schema mapping into a sparse semantic encoding, so any decoder (a renderer, a simulator, an agent, a generator, a game engine, a neural renderer) can render, replay, or build from the same bus.
Each source ships a declarative mapping. Each mapping emits sparse deltas keyed by entity path × timeline × ontology component. The wire format is plural; the shape is one. Below: a live transcript of the mapping claude-session-history-v1 ingesting ~/.claude/session-stats/history.jsonl into XRAI SSE at ~4 Hz. This is running right now in your browser.
The master ontology is not a tree. Every component, archetype, and relation is tagged along eighteen independent axes. Decoders filter on axis tuples: a physics simulator subscribes to A13 = classical-newtonian; a cultural historian subscribes to A15 = {social, cultural}; a diffusion renderer subscribes to A9 = latent. The ontology learns — unknown slugs arrive from ingestion, align via embeddings, promote via review.
A stock tick, a motion-capture joint, a gene expression event, and a quantum orbital land on the same bus with the same delta shape. The difference is which facet tuple they populate. Decoders decide what to do with each stratum: render as space, chart as series, simulate as physics, narrate as story.
A decoder declares what ontology components it reads and what it produces — pixels, a simulated world, a downstream stream, a converted file, tokens, actions. A single SSE stream drives many decoders in parallel. Agents and generators are first-class: they can emit back into the bus.
HDRP + VFX Graph draws Portals holograms on iPhone + visionOS. 370+ shader networks wired from SSE components at 11 ms latency.
NeRF-on-the-fly / 3D-GS / diffusion-conditioned renderers pick up any pose + latent stream and render photoreal without re-authoring.
Reads a seed entity + voice intent, writes new SSE deltas. “a mystical forest with floating lights” becomes a subtree of ontology-typed entities.
Tool calls and chain-of-thought roll onto the bus as deltas. jARvis consumes scene state, emits edits, runs inside Portals or any MCP host.
Replay an XRAI session in the Rerun viewer, or decode an existing .rrd as SSE via the inverse mapping. Zero lock-in.
This page is a decoder. It reads xrai://graph/self, renders the hypergraph, scrubs the timeline. Try shift-click a node.
Any XRAI document must display somewhere. Twenty-three tiers of fallback, from 1-bit existence to multiverse simulation. The host negotiates the highest tier it can render given the installed decoders + bandwidth + power. Every runtime ships T0–T6 as built-ins — worst case any doc is still readable as ASCII + emoji + a color-coded graph. HTTP works on 2G and on fiber; XRAI works on a terminal and on a render farm. Below: the same entity /scene/tree_1 displayed at every tier.
fn frag() {
col = sdBox(p, r);
return ring(col);
}
Every decoder declares produces.tier in its manifest — the host negotiates the best match. Down-tier projection is lossy-but-readable (T22 universe → T6 node graph → T4 🌳). Up-tier synthesis is optional and tagged with provenance.inferred=true so downstream viewers can dim or flag guessed content. T0–T6 are mandatory built-ins.
Anything that would bump the version touches an RFC. Discussion first, implementation after, rejected RFCs archived with a postscript. No silent drift.
Each of these is an XRAI decoder consuming the same SSE bus. Open all three in tabs; they read the same documents. This very page is one of them — the force-graph + html/css decoder rendering xrai://graph/self.
XRAI stores rules, seeds, and intents — not pixels, not geometry, not the full weight of the world. The software on the edge and in the cloud endows it with its look, dynamics, and interactions. The format itself is as lightweight and generative as the laws of physics.