.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
ProofLoop: Evidence-Backed AI Coding
AI-assisted software development can produce code quickly, but speed does not guarantee correctness. A generated implementation may satisfy the visible feature request while violating an important business invariant that was never explicitly tested. Developers therefore need a way to move from “AI generated code” to “AI-assisted code with evidence.” ProofLoop is an evidence-backed developer workflow designed to solve this problem. It wraps an AI-assisted code change in a structured verification loop: Intent → Change Contract → Implementation → Adversarial Challenge → Repair → Deterministic Verification → Proof Pack. Our hero scenario is promotional coupon support for a checkout system. The workflow defines a critical invariant: the payment amount must never be negative. An adversarial analysis deliberately challenges the implementation with a $50 subtotal and a $75 coupon, exposing a potential -$25 payment. The defect is repaired with a zero-floor guard and a regression test. The implementation is then checked using real pytest, mypy, and ruff execution. The resulting evidence is assembled into a traceable Proof Pack that reaches VERIFIED only after the required checks succeed. The target users are developers and teams using AI coding assistants who want stronger confidence in AI-generated changes without turning development into a manual review process. Developers interact with ProofLoop through the workflow and its dashboard, which exposes the contract, adversarial finding, repair, deterministic verification results, and final evidence artifact. What makes ProofLoop different is that it treats verification evidence as a first-class development artifact. It distinguishes LLM reasoning from deterministic tool evidence instead of presenting both as equivalent. The adversarial step actively searches for invariant violations rather than only checking whether the requested feature exists. The result is a workflow that does not simply ask whether AI produced code. It asks whether the change can produce evidence that it is ready.
27 Sep 2026
