Natural-language legislation

Policy mechanics have been menus: a fixed list of toggles chosen by designers. Constrained decoding can compile free-text proposals into a typed policy object over the simulation's actual parameters, and language-model interest groups can respond in character with positions grounded in how the policy affects their simulated welfare. Player-written law moves governance games from choosing among authored options toward authoring rules, which A Tale in the Desert and Eco explored with players as legislators and humans as the parsers.

The player drafts 'Ban coal plants within 2 km of housing; subsidize heat pumps by 30 percent.' The compiler returns a structured policy card listing the parameters it maps to and anything it could not map. Interest groups (miners, residents, industrialists) post positions computed from projected welfare changes and voiced by a model. Votes and compliance follow simulation state, not rhetoric.

Build a policy compiler over a 40-parameter city model. Collect 300 player-written laws, have the compiler map them, and have two humans judge mapping fidelity. Success: 85 percent of laws mapped faithfully or with clearly flagged gaps, and playtesters prefer authoring to menus in a within-subjects test.

• Players write laws the simulation cannot represent; the unmapped remainder must be visible.
• Group rhetoric can diverge from the computed welfare that actually drives votes.
• Politically sensitive content needs moderation.