BountyProof

What it does

BountyProof is a prototype that helps developers inspect public software bounties before spending time on them. It collects a bounded public GitHub evidence snapshot and reviews availability, payment conditions and eligibility with source-linked findings. Exact-quote checks reject fabricated citations, while errors and unknown conditions remain visible. The included deterministic replay exercises the real Mozaik runtime; it is not a live AI inference demo. Six automated tests pass. A configurable model adapter is included, but local-model experiments did not establish reliable live review. Built with AI assistance.

How the agents run concurrently

Three Mozaik agents start independent runLoop calls on the same review.requested event. Availability, payment and eligibility findings update a shared RuntimeState evidence ledger with a revision counter. A fourth critic starts as soon as two findings arrive, potentially while the third specialist is still running. If a late finding changes the revision during the critic review, the stale result is discarded and the critic runs again. An observer validates source quotes and records the event trace. Reproducible replay tests verify overlapping specialist execution and stale-review invalidation; inference responses in that replay are explicitly scripted.