# sqlswarm: a fair head-to-head for coding agents

An experiment setup for answering one question honestly: does a swarm of Claude Code agents inside throng beat one plain Claude Code session on the same hard task? The task is writing a SQL engine in Rust from a 20-line stub, graded by SQLite's own sqllogictest suite.

The design is the part worth copying. The grader, tests and corpus are written before any agent starts and are read-only to the agents; scores are always computed from the untouched scaffold plus the agent's src/, so an agent that edits tests cannot raise its score.

## Files
```
PROOF.md               the claim, who wrote what, how scoring works, how to check every number
scaffold/goal.txt      the exact goal both arms get
scaffold/arm_solo.py   Arm A runner: headless Claude Code for a fixed wall clock, graded from outside every 5 minutes
scaffold/NOTES.md      every change to a non-src/ path, with its reason
proof/baseline.json    the stub's score: 30 of 11,185 records
proof/oracle.json      real SQLite on the same grader: 11,183 of 11,185
```

The grader (slt/), generated tests and the Rust score example are not included here; they are Rust source.