concept (lookup) ·
concept rel (apply rotor) ·
A : B :: C (analogy) ·
concept → rel → rel (chain) ·
what is X (profile)
·
natural language: any free-form sentence — auto-resolved against the codebook via the grounded LM, with per-token anomaly scan
03·2027
This is a deterministic symbolic reasoner. The codebook holds 137,199 entities (WordNet synsets + ATOMIC commonsense phrases) embedded as vectors in the Cl(3,3) bivector algebra. Relations like _hypernym, _part_meronym, _causes, xIntent, Causes are stored as rotors in the algebra — fixed transformations across all 38 relations.
When you ask tree _part_meronym, the machine: (1) finds tree.n.01 in the codebook; (2) applies the _part_meronym rotor — a precise algebraic rotation; (3) finds the nearest codebook entries to the result. Output: burl, trunk, crown, stump, limb — exact, every time, no temperature.
Analogies work the same way: king : queen :: man derives the rotor that takes king→queen via the wedge product, then applies it to man. The result is the position in primitive space that the analogy predicts.
◢ Grounded disambiguation (Path B): Natural language doesn't get translated into a symbolic query. The language model's hidden states are bound to the same Cl(3,3) algebra the symbolic engine operates in: at convergence the v2 LM holds substrate-binding loss to ~2 × 10⁻⁴ per token — the chain's per-stage hidden state matches the algebraic prediction to that precision. The neural parse is already in the algebraic substrate before the symbolic engine touches it. No translation layer, no interpretation step, no semantic gap — this is what the substrate-binding research makes possible.