SkyNet

VerifForge — SkyNet AI can suggest a useful engineering experiment. The harder question is what actually ran, what was observed, and which conclusion those observations support. VerifForge is a local engineering investigation engine built around typed proposals, bounded execution, independent immutable checks and evidence with source, stimulus, checker, contract and environment identities. The retained-evidence demonstration shows one accepted Meta-to-tool feedback chain. An actual muse-spark-1.3 response proposed a 17-cycle FIFO experiment. That experiment ran in the identified hosted native job and received the checker verdict pass_within_scope. The observations were imported, checked again and persisted as feedback. The original bounded followup expired without invoking the provider and was later cancelled. A separately authorized actual Meta request consumed the saved feedback and returned a valid 13-cycle proposal. That second proposal was not executed. The first experiment reset an initially empty FIFO; it did not expose the known reset defect and is not a bug-discovery claim. A separate recorded firmware example makes the distinction between process completion and design verdict concrete. A deliberately faulty header writes STATUS at the wrong register address. The native MMIO checker reports the first conflict at cycle 95. A curated, hand-authored one-line correction passes under the unchanged checker and contract, and saved-program replays preserve the bad and candidate observations. This is a separate configuration from the Meta/FIFO chain; Meta did not generate that correction. Our event-period contribution is the integration behind the recorded Meta/FIFO feedback chain and firmware/MMIO check/replay path, demonstrated on 4 November and 30 October. The earlier deterministic FIFO/platform foundation predates the event. PicoRV32 and the execution tools remain upstream work. The demonstration makes both the proposal-to-feedback chain and the firmware mismatch inspectable through retained evidence. Browser qualification, wider recovery, clean final-release reproduction and the full evaluation remain open. Public source and selected sanitized evidence are available; no hosted product deployment or physical-hardware validation is claimed.