one dataset · paper, tables, graph, wiki
simgrammar the grammar of simulation

core loop

start with water, fire, wind, and earth -> drag two elements together -> look up the pair in the shared database -> if new, a language model names the result and picks an emoji -> result is cached for everyone -> combine the new element further

design notes

At runtime the model is consulted only on a cache miss: the game checks a global database for the unordered pair, and if no result exists, Llama 2 (later Llama 3.1) generates a short element name, capped at twenty tokens, plus an emoji; the result is stored so the same pair returns the same element to every player. The recipe graph is therefore written collectively and lazily by players and the model, and a First Discovery label marks players who triggered a generation. Pairwise combination discovery is not yet a mechanic in the taxonomy (proposal combination-discovery). It is a boundary case: there is no simulated system beyond the generated recipe graph.

relations

3 stored edges
uses

Combination discovery

Every action combines two owned elements; on a cache miss a Llama model names the result, the result is stored so every player gets the same answer, and new results earn a First Discovery label.

core
uses

Generative content on demand

On a cache miss, a Llama model generates the combined element's name (up to twenty tokens) and emoji; the result is written to a shared database so later requests are deterministic.

core
uses

Collection goals

Players accumulate discovered elements and First Discoveries, which mark elements no player had made before.

supporting

backlinks

computed from other records

sources

1
Wikipedia contributors (2026). Infinite Craft. Wikipedia

provenance

runs in log/runs
2026-10-24
created2026-10-24-catalog-gdcatalog