symbolic mind
An object neural network built from scratch — its substrate is objects, not tokens. 70 GB of source code translated into layered object data across many levels of abstraction, with the finite types of mechanics and algorithms formalized at every level.
Objects, not tokens.
Today's networks learn distributions over tokens; what they know is trapped in weights. symbolic mind is built on a different substrate: typed objects and the relations between them. 70 GB of source code has been translated into object data layers at multiple levels of abstraction — so what the network learns can be read, validated and diffed, not just sampled.
Formalized structure at every level
Objects at the bottom; mechanics above them; algorithms above those; systems at the top. Each abstraction level carries its own formalized vocabulary — so what the network learns can be validated and diffed rather than sampled. The exact representation is withheld pending IP review.
Knowledge is the force of gravity.
In this architecture accumulated knowledge acts as an attractive force in the formation of structure: the denser the knowledge around a thing, the stronger it pulls new structure toward it. Graphs are not assembled by rule alone — they condense around mass, the way matter condenses in a gravitational field. What the network already understands shapes what it can learn next.
Meaning graphs of working discussions.
The first instrument built on this discipline turns the transcript of a working discussion into a graph of meaning: who acts, what connects to what, which rules apply — and, critically, where participants fill the same word with different meaning. "Order confirmed" means a signed specification to the commercial director, a checked box in the CRM to the manager, and a payment fact to the database. All three are right; the argument lasts weeks because the word is one and the meanings are three.
- A closed vocabulary, fixed in advance and validated: anything outside it is rejected rather than accommodated, which is what makes two runs comparable.
- Divergences are flagged, not smoothed over: each interpretation is kept with its author's own words, and the conflict is recorded rather than resolved silently.
- Abstraction is derived rather than stored, so the hierarchy comes out the same on every run.
- Human edits are protected: what a person confirmed is never overwritten by the next run — a difference becomes a proposal, and the human decides.
- Quality is measured against a fixed set of expectations, including the generalizations that must not appear.