The dream of an AI game master is old and the failures are familiar: the model contradicts its own rules, loses track of state, and railroads the player to keep the story on rails. AI Dungeon made this legible to everyone — generative, surprising, and incoherent over any real horizon.
The instinct is to reach for a bigger model. That's the wrong axis. Reliability over a long session is not a capability problem; it's an architecture problem.
Separate what can be decided exactly from what needs judgment
A rulebook contains two very different kinds of statements. Most rules are determinate — given the state, the outcome is exact and checkable. A few are genuinely ambiguous — they need interpretation, taste, or narrative judgment.
A pure-LLM GM treats both the same way, so the determinate parts inherit the model's drift and its un-auditability. The fix is to split them:
- Compile the rulebook into a constraint engine. Parse any rulebook into a probabilistic rule graph of rules and their conditions — a queryable artifact, not prose in a context window.
- Route determinate rules to a symbolic checker. These resolve deterministically: 100% accurate, no model in the loop, no drift, fully auditable.
- Send only the ambiguous remainder to a grounded LLM arbitrator — one that reasons against the compiled rules rather than inventing them.
The result is a GM that can run for hours without breaking its own rules, where every ruling traces back to either the rule that produced it or the arbitration that resolved it.
The same pattern recurs
This neuro-symbolic split — LLMs propose, a deterministic engine enforces — shows up across the strongest agentic-game ideas:
- Emergent norms in NPC societies: let the models decide what norms exist, and a signed, append-only commitment ledger decide what norms enforce. Prior work conflated the two; separating them gives you societies that are both emergent and binding.
- Persistent world-state: the soft narrative layer (embedding-retrieved memories, kept for personality) is split from a hard predicate layer the engine owns as ground truth, with a contradiction checker gating retrieval so the NPC never speaks about a king the player watched die.
- Drama pacing: steer narrative tension from the response side against a learned reward model — never constraining player input — to resolve the agency-versus-coherence paradox without railroading.
The throughline is the same as the rest of the agentic frontier: don't ask the model to be reliable about things that can be made exact. Make them exact, and spend the model's judgment only where judgment is actually required.