Trojan PR

A PR check for ML training code that executes the math instead of reading the diff, in gVisor sandboxes on Vultr, behind NetBird with zero open ports.

Live demoGitHub0

Overview

I've merged more than 140 fixes into ML libraries like DeepSpeed, Unsloth and sentence-transformers, and about two-thirds were silent math bugs: no crash, CI green, approved by human and AI reviewers, and a wrong number for months or years. One DeepSpeed bug returned 53 instead of 9 for 933 days. Trojan PR is the missing layer in PR review: it checks the math, not the diff. A model on Vultr Serverless Inference reads each PR and picks math rules (invariants) the code must obey, and the rules differ per PR. Those rules run in gVisor sandboxes on a separate Vultr VM, on the code before and after the PR. A deterministic judge blocks only if a rule held before and breaks on the PR, and signs an ed25519 receipt. On 12 real bug-causing PRs it raised zero false alarms on the 12 fixes, while an AI reviewer flagged 1 to 2 correct fixes. Everything is served through NetBird with a PIN and zero open ports.

Note: the demo URL is substituted with the GitHub repository link.

Tech stack

No comments yet

Be the first to comment on this project.