🤓 Latest Submissions
Proofline: Trust Layer for AI Coding Agents
The problem: AI coding agents like IBM Bob can now write and change code faster than any human can carefully review it. That creates a gap. Either someone rubber-stamps changes they never really checked, or the agent's speed goes to waste while every change gets the same slow, manual review. Proofline is a verification layer that sits between an agent and the codebase it wants to change. It scores every module's risk (GREEN, YELLOW, RED) from AST complexity, test coverage, and git history, and prints the exact rule behind each score. It then gates what the agent may do based on that color: GREEN allows delete or refactor, YELLOW requires verification to pass first, RED blocks all automated changes and demands a human. Before any change is applied, an adversary step reruns the real test suite and a generated set of edge cases both before and after the change, comparing every single result. Only if nothing changed does the edit get applied, on its own git branch, with a certificate. Every step, human or agent, is written to a hash-chained, tamper-evident audit log: edit one entry afterward and the chain visibly breaks instead of silently staying green. Target users are engineering and platform teams adopting agentic coding tools at scale, who need a policy layer that decides what an agent may touch autonomously versus what needs a human, plus an audit trail for teams that answer to compliance. They interact through a CLI or a web dashboard: browse the autonomy map, submit a proposed change, watch a live verdict, and review the resulting certificate. Most tooling around coding agents focuses on generating more code. Proofline focuses on the opposite side: an adversarial process whose only job is to try to disprove a change, and a tamper-evident ledger borrowed from distributed-ledger ideas, applied to something mundane and useful, an audit trail for AI-authored changes. We proved it end to end on Brightshop, a deliberately messy demo shop app with real dead code, hidden business rules, and uneven test coverage. Proofline correctly blocked an unsafe deletion on a high-risk module, and safely verified and applied a real refactor on a lower-risk one, generating a full certificate from real data, not a mockup.
27 Sep 2026
