Êi|gen⟩ ⊗ Ên|gen⟩
An operator that turns a corpus into a converged, evidenced artifact with an auditable trace. This page is one of its artifacts.
The falqon manuscript's concept graph — 91 nodes, 140 edges, four clusters. The rendering needs JavaScript; the same object is walkable in the demo below.
Four clusters, sized by how many edges each node carries; the section spine runs down the left. Hover a node for its full label, id and edges — on a phone, tap.
Six def nodes ring in red: definitions no claim consumed, caught by the connectivity gate and shipped as they are.
Today's models forget you between sessions, and left to themselves write from the mean of everything they have read. These are the same failure. The machine runs on one enormous forward map, steered after the fact by prompts and rewards - a wall you lean on, never a generator that makes. EigenEngine runs the arrow the other way: it builds a person's knowledge into an explicit graded structure, the conceptric, so that memory becomes reconstruction rather than storage and generation becomes projection rather than search. A small, local model routes a question down that structure and spends a large model only on a brief that already half-knows the answer - the corpus never leaving your own hardware.
It is a ladder of loops, each one useless until it is bootstrapped shut. The architecture is not an analogy borrowed from physics: it is the graded decomposition of the synthetic preprints, made into something you can switch on. The end of the climb is plain - a way for anyone to grow a partner intelligence out of their own life and work, a mind that is theirs, local, that learns them, and answers as a colleague rather than a service.
The code stays private. It holds a corpus you could reconstruct, and the mathematics that built it is the other half of the preprints. What is public is the output and the demonstration - and this introduction was written, start to finish, by the engine it describes.
One staged operator, drawn as the stages it has. Hover a stage for what it does; the projections fan from the same converged object.
The sources are decomposed into a graded graph of typed nodes and relations — 91 nodes, 140 edges in the falqon build shown above.
conceptric_v1.json · 91n/140eThe paper and the talk are the same material projected into two forms, a manuscript and a set of slides, each generated by the engine and each pointing to the other.
A physics manuscript built through the full pipeline: claims curated with every number keyed to its numerics readings file, the concept graph converged under gates, the residue tracked instead of smoothed. The demo is the engine's own editor over that manuscript's real conceptric — the object rendered at the top of this page.
The Spanspermia deck was generated by the engine from the spoken opening of an FQXi discussion. The Laws of Learning ran the other experiment: v1 the first agent-drafted cut, v2 the hand-tuned pass, v3 the author-fine-tuned deck that survived — the lineage is kept, not hidden.
Each one is enforced by code, not policy. The mono line names where.
Whatever writes an artifact does not get to score it. Manuscript claims are judged by separate, blinded passes; the drain that consumes findings validates evidence before it counts anything closed.
dispatch/judge separation · blind panel protocolNo worker is handed a pile of files. Every dispatched agent is pointed at the structure and told what to reconstruct and for what goal; it builds its own minimal frame and reports back up. Flat reads are the failure mode the architecture exists to remove.
compile_dispatch · reconstruct(node, goal)Nothing counts as resolved without an evidence-carrying record — including the orchestrator's own claims. A gap is named and stays open rather than being smoothed.
friction ledger · drain validatorDispatched workers run at or below the orchestrator's own model tier, chosen by the grade of the work; an audit enforces the rule over every workflow rather than trusting the author of the dispatch.
audit_dispatch_models.pyA system that only exhibits successes is advertising. These are real refusals from the engine's own record, kept because they are what the gates are for.
The first bake of this site's demo bundle was rejected by the engine's own leak scan — an author-identifying literal had ridden into a shared asset. The build shipped only after the patch that stripped it.
site_seal scan · 2026-07-11Three findings were claimed closed in a commit message; the consumer refused to count them consumed until evidence-carrying resolutions were banked. The machinery does not take its own operator's word.
friction drain · evidence-or-openA connectivity gate on a manuscript's concept graph caught definitions no claim consumed. The state shipped as it was — the six red-ringed nodes above — rather than being wired over.
connectivity gate · def-orphansFeedback arriving from the public demo is minted to the visitor who wrote it. Records that arrive actorless, or dressed as someone they are not, are refused and kept for inspection instead of being ingested.
remote-channel actor guardThis is a static page on a static host, so every claim of liveness is bounded the same way: baked at commit. The shell around this text — head, navigation, footer, identity tokens — is generated by the engine's site backend from a spine file; the content region is authored innards; a drift gate fails the build if the generated shell and its sources disagree. The graph at the top is the falqon manuscript's real conceptric, extracted from conceptric_v1.json at the engine tree on 2026-07-11 and drawn client-side — no data leaves the browser.
The content was written by the engine's own orchestrating self during the session of 2026-07-11, from the engine's banked process records — session handoffs, friction ledgers, gate verdicts. The trace shown on this page is curated for reading; the underlying records are the engine's working files, and every exhibit above names its source in the mono line it carries. Where the engine's next milestone lands, this page's content region is scheduled to be compiled from the engine's own concept graph through the same pipeline as its manuscripts — at which point this fold will say so, with the run's own record.
shell: site_backend · spine: site/spine.json · innards: site/content/eigenengine.html · graph: assets/js/eigenengine-graph-data.js · gates: drift · links · survival · meta · deslop · visual(375px)
The demo is the engine's conceptric editor over a real manuscript corpus, running entirely in your browser. Walk the graph, open a node's essence, follow an edge to its evidence. If you leave a comment, it becomes a typed record and lands in the engine's intake through the same gates as everything else on this page.