Agent-prepared evidence for human review. Read data__RESULTS.md for the complete16-cell table and limitations. The sparse count-aware observations are not a stable general multiplier; small/default/some domain controls are mixed or slower.
Download this gist as ZIP, extract it, then run python3 restore_bundle.py. It creates a fresh ready-status-evidence-20261103 directory and verifies the SHA256 of all 40 files. Read that directory's README for the exact beads-only setup, actual executed harness/runner, source snapshots, commands and raw logs. No Gas City installation or production database is required. Use a disposable environment.
The captured experiment uses Dolt 2.6.0/Go 1.27.3, one CPU quota/6 GiB, two batches per arm and448 measured operations. It measures the specified SQL-backed operation with transaction boundaries, not deployed CLI/fleet latency. Both arms use indexes; the follow-up tightens status ranges. Original subqueries are cacheable.
Source