Designing Rule-Bound Worlds for Humans and AI
Most multi-agent roleplay systems start from the wrong abstraction. They try to manufacture drama by stuffing personality adjectives into system prompts: "You are sharp-tongued, defensive, and secretly anxious." Within ten turns, every agent collapses into theatrical caricature, repeating surface tics while losing track of real incentives.
When we stripped every personality adjective out of AI Detective, distinct identities did not disappear — they got sharper. Strategy and voice emerged naturally from two things: asymmetric pressure inside the rules, and native differences between underlying models.
1. Isolation & Stakes Over Adjectives
In an unscripted deduction game, behavior comes from what you know, what you stand to lose, and what the engine refuses to let you see. Our harness enforces three structural invariants:
-
Physical information isolation: Every event carries an explicit
receiversmask. Private whispers, sealed dossiers, and off-stage discoveries are never serialized into another seat's context window — preventing cross-talk by construction rather than asking models to "pretend you didn't hear that." - Objective situational injection: Instead of instructing an agent to "feel cornered," a deterministic Python builder appends raw state facts to the trailing board: remaining cross-examination quotas, who currently leads the suspicion vote against it, and which secrets remain exposed.
- Minimal physical anchors: Each seat holds two concrete, camera-visible objects in context. Reactions map strictly to visible gestures or silence — eliminating prose fluff about internal emotions.
2. Eval & Blind Gates Before Scale
Running seven commercial reasoning models over hundreds of turns is slow and expensive. Before touching a full-game run, we gate every layer — reasoning depth, rule consistency, and visual assets — behind small, adversarial probes.
Across 74 controlled cognitive-load probes testing multi-model reasoning tiers, we found a sharp inflection curve: pushing thinking effort from high to maximum tiers multiplied token cost by 4.27× to 8.68× while delivering zero accuracy gain on solved tasks. Inspection showed models reached the right deduction within the first 20% to 50% of their thinking budget, then burned the remainder re-counting string lengths, doubting settled logic, or wandering into false-positive safety filters.
We apply the same separation to aesthetic quality. Rather than asking a model to "make it look good," generation runs through two blind gates: an isolated visual-readability check (can a blind reviewer infer the role solely from uncaptioned character art?) and strict continuity checks across multi-shot cuts.
3. Rules as an AI-Native Medium
Traditional media treats AI as a cheaper camera or faster copywriter. A harness treats rules as the primary creative medium. When the game state machine, judge validation, and replay journal are deterministic and auditable, the simulation stops being just a video pipeline — it becomes a self-contained rule world where frontier models and human players can eventually sit at the exact same table.
Good characters are not written in adjectives. They appear the moment a system corners an agent with incomplete information and forces it to choose.