Model Selection Guide

Geopol Forecaster uses OpenRouter as a unified API gateway, giving access to models from every major provider through a single endpoint.

Model Pools

Each pool defines 4 roles:

Role	What it does	Quality needs
Planner	Runs once — web research + briefing synthesis	High reasoning
Narrator	Adjudicates rounds, weaves outcomes	Strong creative + analytical
Player	Responds in-character each round	Good persona adherence
Advisor	Supports human players with analysis	Basic competence, fast

See config/pools.yaml for all predefined pools. Set active_pool or pass the pool name at runtime.

Browse Models

• OpenRouter Roleplay Models — browse models tagged for roleplay
• OpenRouter Rankings — overall model rankings by category

Roleplay Benchmarks

Use these to track which models perform best for character simulation and creative interaction:

Benchmark	What it measures	Link
EQ-Bench	Emotional intelligence, empathy, social skills in roleplay scenarios	eqbench.com
LLM Stats Roleplay	Aggregated roleplay benchmark scores across models	llm-stats.com/benchmarks/category/roleplay
PingPong Bench	Dedicated RP benchmark using LLM-as-judge evaluation	github.com/IlyaGusev/ping_pong_bench
RoleLLM / RoleBench	Character-level roleplay benchmark (168k samples)	github.com/InteractiveNLP-Team/RoleLLM-public
Arena AI	Crowdsourced side-by-side model comparison	arena.ai/leaderboard

Adding Custom Pools

Edit config/pools.yaml to add your own pool. Use any model ID from OpenRouter's model list:

pools:
  my-custom:
    planner:  "provider/model-name"
    narrator: "provider/model-name"
    player:   "provider/model-name"
    advisor:  "provider/model-name"

Set active_pool: my-custom or pass it as a CLI argument.