The pipeline

Every step assumes contributors might lie. The swarm proposes; infrastructure we control decides what counts.

From source to specimen

A maintainer opts a project in. We build the target from source with coverage instrumentation and sign the result, so workers only ever run code we compiled. The swarm mutates inputs from a shared corpus, and every worker starts from what the swarm has already discovered rather than from scratch.

Scoring, not raw counting

The coordinator merges new coverage into the corpus and discards duplicates. Re-running work the swarm has already done earns nothing. This is deliberate: it stops anyone with a warehouse of machines from farming rewards by repeating worthless tests.

Reproduce before you trust

A crash only counts after independent workers reproduce it in an environment we control. Contributor telemetry is never trusted on its own. Once a bug is fixed, it becomes a permanent public specimen with its whole history attached.