A character should remember who they are. And notice what has changed.
Ask the same NPC something in a private room, after a betrayal, or in front of an enemy. The voice should still feel like theirs. The answer should not stay frozen.
NPCBank makes each dialogue case inspectable: the prompt, the character's current state, the source evidence behind it, and the rules used to judge the response.
Corpus
- Universes
- —
- Characters
- —
- Cases
- —
Grounding
- Source documents
- —
- Extracted entities
- —
- Tracked state variables
- —
Drawing an entry from the store…
One of 9,785 cases, drawn live. The margin holds the condition the case is asked under. Switch the view to see what the schema keeps out of the model's reach.
Where the cases fall
Counted from the store, now.
How a case is compiled
-
The source
A document keeps its structure, its speakers, and its quoted speech. What can be lifted out of it (a claim, an event, a relationship, a moment in the timeline) is lifted with a pointer back to the line it came from. Cases are compiled from that evidence, which is what lets the collection extend to new games without proportional annotation effort.
-
The character
What a character is stays apart from what is currently true of it. A profession is identity; a broken trust is state. Collapsing the two is how a benchmark stops being able to say why a model behaved differently.
-
The case
One question, one capability, one condition to answer under, plus the invariants that hold whichever condition applies. Change is meant to be read by asking the same question twice with one condition altered and every other difference recorded.
-
The contract
Required evidence, observable signals of success and failure, the response mode, and the failure families an error is attributed to. Written before the model runs, and held on the judge's side of the schema.