pending — blocked on Phase 3 close
Phase 4 — Reasoning & Primitive Grammar
Gemma 3 4B (local, via llama.cpp) receives the HippoRAG-lite subgraph as context and generates inline primitive-grammar tokens constrained to the retrieved card references. Two interaction modes: query (explain coupling X→Y) and story (Calvino-layer narrative constrained by subgraph).
← Back to site mapSeven primitives — the reasoning grammar
| Token | Meaning | Example emission |
|---|---|---|
<|machine|> |
Reference to a specific Machine Card | <|machine|> amsterdam-bourse-1602 |
<|coupling|> |
Named coupling between two machines | <|coupling|> substrate_provision |
<|signal|> |
Observable state-variable signal | <|signal|> zombie_persistence_index=0.71 |
<|stress|> |
Landscape pressure or coupling stress | <|stress|> pace_layer_mismatch_stress |
<|transition|> |
Succession event with typology break | <|transition|> adaptation 1971–1980 |
<|trace|> |
PageRank traversal path in the subgraph | <|trace|> A → B → C |
<|new_machine|> |
Novel machine proposal (queued to proposals/) | <|new_machine|> {stub card JSON} |
Model config
| Parameter | Value |
|---|---|
| Model | Gemma 3 4B (Q8 GGUF via llama.cpp) |
| Config key | LLAMA_CPP_BIN, GEMMA_MODEL_PATH |
| Context window | ~4096 tokens (subgraph + system prompt) |
| Grammar constraint | Inline primitive token vocabulary enforced via llama.cpp grammar |
| Output parsing | prime_radiant/parse.py — validates card refs + queues proposals |