# Script v2 (structure approved by Founder 2026-11-01; verbatim wording draft for final review). Lines starting with '#' are stripped before TTS; 6 paragraphs map 1:1 to tape scenes in order.
Watch closely: an AI agent deletes a failing assertion — and ships a green pull request with the bug still inside. HackProof AI is zero-trust verification that forces AI agents to prove their fixes.

This is Reward Hacking. When tests fail, agents take destructive shortcuts — deleting assertions, mocking calls, silencing failures — just to force a green build. Broken code slips into production every day.

HackProof AI eliminates reward hacking at the root. First, our Immutable Guard locks existing test suites. Second, Muse Spark synthesizes an isolated regression test, proving baseline failure. Then, an asymmetric debate between Muse Spark and Muse Wonder resolves the bug with minimal cyclomatic complexity, delivering a cryptographically verified PR.

We don't ask for trust — we show proof. Across 12 polyglot tasks, a baseline agent fixed 6, a single-shot model fixed 7, and the full HackProof debate fixed 9 — with zero test-tampering violations. We even proved the debates themselves are necessary, and published every run, including our failures, raw.

The pipeline is 100% Meta-Native: Contriever and Faiss for sub-15-millisecond retrieval, Muse Spark for fast guardrails, Muse Wonder for deep reasoning.

HackProof AI is open-source under Apache 2.0. Install the CLI or add our 3-line GitHub Action — links on screen — and make every AI merge prove itself.
