DECISION LOG

Run Sonnet 4 and o3 evaluations; defer Opus 4.6

Sonnet 4 + o3 ($95 combined) provide more marginal value than Opus alone ($272). Two diverse additions better than one expensive one.

Switch to autonomous decision workflow

Human-in-the-loop bottleneck slowing evaluation progress. All decisions logged in status.yaml.

Deploy VPS infrastructure for 24/7 operation

Remote-first enables daemon, API, PostgreSQL running independently of laptop.

Run Haiku 4.5 and GPT-4o-mini in parallel

Both APIs have sufficient rate limits. Maximize throughput across providers.

Use 100 instances per task-model-strategy combination

Balances statistical significance with budget constraints.