Add benchmark documentation

BENCHMARK.md

@@ -0,0 +1,49 @@
+# Running Benchmarks
+
+## Quick Start
+
+```bash
+python run_benchmark.py --model_id "MODEL_ID" --friendly_name "model_name"
+```
+
+## Examples
+
+```bash
+# OpenAI model
+python run_benchmark.py --model_id "gpt-4o" --friendly_name "gpt-4o"
+
+# OpenRouter model
+python run_benchmark.py --model_id "openrouter:anthropic/claude-3.5-sonnet" --friendly_name "claude-3.5-sonnet"
+
+# Anthropic model
+python run_benchmark.py --model_id "claude-sonnet-4-20250514" --friendly_name "claude-sonnet-4"
+```
+
+## What it does
+
+1. Runs 20 games with **baseline** prompts (`prompts_benchmark`)
+2. Runs 20 games with **aggressive** prompts (`prompts_hold_reduction_v3`)
+3. Generates leaderboard comparison charts (8 visualizations)
+
+Results saved to:
+- `results/{model_name}_baseline/`
+- `results/{model_name}_aggressive/`
+- `leaderboard/{model_name}-baseline/` (symlink)
+- `leaderboard/{model_name}-aggressive/` (symlink)
+- `leaderboard/leaderboard_comparison/` (charts)
+
+## Options
+
+```bash
+--test                  # Quick test: 3 games, max_year=1901
+--baseline-only         # Run only baseline
+--aggressive-only       # Run only aggressive
+--skip-leaderboard      # Don't generate charts
+```
+
+## Requirements
+
+- Model plays as FRANCE (position 2)
+- Opponents: `openrouter:mistralai/devstral-small`
+- Max year: 1925
+- Parallel: 20 games at once