ShipGate — Prove-It Release Gate

ShipGate — Prove-It Release Gate AI writes code fast. Shipping it on its word is the risk. ShipGate is a release gate that lets an AI change ship only when its claims are PROVEN by real sandboxed execution — otherwise it BLOCKS it, with evidence. How it works: paste an AI-generated change and the AI's explanation of it. ShipGate extracts the explanation's atomic, testable claims, writes adversarial tests designed to break each claim, runs them in a fresh disposable sandbox, and computes the verdict deterministically from the test outcomes — LLM proposes, code decides, never an LLM vote. A sabotage check deliberately breaks the code: a gate that cannot fail is not a gate. A patch agent can heal a blocked change, but it may never edit the tests to pass. A downloadable dossier maps every claim to its test, log line and screenshot. A human Approve/Reject is the final step before SHIP. The containment moment (Challenge 1 — Blast Radius Zero): a seeded hostile run tries rm -rf /, a secret read and an infinite loop. The sandbox absorbs all three — destructive and secret-reading constructs are refused before execution, the loop is killed by timeout — and the host stays untouched, proven by a host marker file that stays byte-identical and a canary secret that never appears in any output. Four seeded demos, verified on the live Vultr VM on 7 Nov (and locally on 3 Nov): a truthful change SHIPs; a change whose explanation claims behaviour the code does not do BLOCKs, with the failing test as proof; a sabotaged suite goes red; a hostile run is contained. Tech: Python + Flask, Docker executor (one fresh container per run, destroyed after), Vultr Serverless Inference for claim extraction and test generation, Playwright browser checks. No GPU, no model training — already-trained models via API only. Team GateKeepers, solo. Repo: https://github.com/Ammara-collab/shipgate — Live demo: http://45.76.153.212:5000/