Reveal the strongest layer-five dimension at one token position. A local clue, not the final destination.
Project // Live Model // Game
Semantic Blackjack
Write the hand. Watch GPT‑2 tokenize it into cards. Stand, and let the model resolve the complete string to one of 3,072 places. The House has two numbers. Language finds the rest.
Table rule // implicit resonance
The House holds two places.
Every hand is a string. Every visible card is a contextually produced GPT‑2 token. Standing measures layer five and sends the complete hand to the dimension with the strongest post-activation anywhere in the sequence.
HOUSE1888 + 1790◆2256PUSH // ANYWHERE ELSE PAYS
CARD 01 // TOKENtheappend-only input
CARD 02 // TOKENseatokenization may shift
CARD 03 // TOKENwasup to eleven cards
CARD 04 // TOKENveryeach position offers a route
STAND // DESTINATION1430the model settles
A model query made playable
Information has a price.
Learn whether either House dimension is present among the complete hand’s global top three.
Hand length trades payout for additional chances to escape the House. Rare destinations and first visits pay more.
One of our mesocosms
Language is the hand, the route, and the wager.
The project turns a frozen internal readout into a small persistent world. Players accumulate discoveries, spent strings, purchased advice, ledgers, rescues, and their own history of arrivals. The model is not pretending to be a dealer. Its deterministic hidden activations are the table.
Two doors // play and inspect