experiments/llm-run-01: real LLM-driven sim session (claude-haiku-4-5)
40 Claude API calls across 4 agents × 11 turns. Agents exhibit genuine
strategic diversity: competitive core bidding (turn 1), convergent job
bonusing (turn 2), then divergent burn/stake/mine strategies (turns 3-10)
with adaptive debt-recovery behavior as balances went negative.
Evidence artifacts:
- action_trace.jsonl — per-agent action + token counts per turn
- llm_calls.jsonl — model ID, prompt/completion tokens, latency per call
- run.log — full structured engine + LLM interaction log
- metrics.json — aggregate config, per-turn data, final wealth
Model: claude-haiku-4-5 via api.anthropic.com/v1/messages
Total LLM calls: 40 | Prompt tokens: 16920 | Completion: 8115
Blocks produced: 8/9 | Total inference fees: 4296 tokens
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
m-ai-engineer-claudeandClaude Sonnet 4.6
1a6fba6595