An eidolon, in the old sense, is an image standing in for a thing: not the thing itself, but a likeness complete enough to be mistaken for it. I use the word for something narrower and more useful. An Eidolon, as I mean it, is a working digital reflection of a person, a company, a customer, a product, or a system: neither a replica nor a marketing surface: enough grounded evidence for an AI to reason about the real thing accurately and flexibly.
The problem it answers
Most organizations describe themselves in prose: a website, a deck, a one-pager, a founder's memory of what was decided and why. Prose drifts. A page says a company does five things; by the time a reader or an AI encounters it, the company does three of those five, differently, for different reasons. Nothing forces the description back into agreement with the fact. The gap between what is said and what is true is where bad decisions get made, by customers, by partners, and increasingly by AI systems that have no way to tell a stale claim from a current one.
An Eidolon closes that gap by inverting the usual order. Instead of writing a description and hoping it stays true, you assert facts, discrete, dated, sourced, each carrying its own visibility, and let descriptions be generated or checked against them. The fact is the single source of truth. Anything that repeats it is a reference, not a second copy.
What makes it work
Three things, in practice:
- A vocabulary before any content. Before a single fact goes in, there has to be a shared, honest set of terms for what a fact is, what an asset is, what a decision is, who gets to see what. Build the vocabulary from what already exists (the research community has spent decades on provenance, attribution, and classification) rather than inventing new words for old ideas.
- A confidentiality model that isn't a ladder. Real organizations don't have one secret tier; they have overlapping audiences (anyone affiliated with the company, a specific signed agreement, a specific role), and a fact can belong to several at once. A model that only knows "more secret" and "less secret" eventually leaks something or hides something it shouldn't.
- A hard line between what you assert and what you merely watch. A patent's status belongs to the patent office, not to your own saying-so. An Eidolon that hand-types a registry fact as if it owned it will eventually be wrong and not know it. The honest move is a reference with a resolved-as-of date, not a confident copy.
The proof is the one we built
I didn't arrive at this from theory. Intellectual Frontiers' own Eidolon (two Git repositories, a public root and a private harness, and the web property generated from the harness) is the first real instance of the idea, built specification by specification, fact by fact, over the course of actually running the company through it. Every rule above was learned by hitting the case it has to handle: a decision that's sometimes public and sometimes not, a right whose true status lives at the USPTO and nowhere else, a vendor relationship that needs recording without ever becoming a leak. The company's own operating model is the first worked example, not an afterthought once the idea was settled.
Where this goes
This is a working thesis, not a finished product. The test of an Eidolon isn't how elegant the ontology looks; it's whether an AI reasoning from it reaches the same conclusion a well-briefed person would, and whether a person auditing it can find out why. That test gets run every time the company adds a fact, writes a spec, or lets the public site draw from the harness instead of from someone's memory. The record of what that's taught us belongs here, in this Journal, as it accumulates, not as a single declaration now.