U1ai

Resources

Evidence over adjectives.

Everything here is public, readable, and labeled with its real status. Reviewers are encouraged to click.

On the record

Research, code, and writing

  • On the public record at NIST

    Submitted a public comment to NIST on the SP 800-82 Rev 4 preliminary draft (Guide to Operational Technology Security), September 2026 - arguing that compliance status is not authority to change OT state.

    Read the submitted comment

  • Open-source authority gate

    Glass Box, our public repository, holds the deterministic Sentinel authority gate (ALLOW / REVIEW / DENY) with a passing regression suite and the Jev evidence-evaluation adapter.

    github.com/Unmute1-Ai/glass-box

  • Published writing

    An ongoing article series on interpretation vs. authority in AI systems, including “PRIMER: Building AI for Children That Knows Its Place.”

    Read on Medium Read on LinkedIn

  • Benchmark in progress Kaggle × DEV challenge

    “Confidence Is Not Authority” - an open benchmark that measures how often models claim authority they were never granted, using frozen adversarial cases (critical false-allow rate). Entry in progress for the October 2026 challenge.

    Read the review package

Try it

Authority gate demonstrator

An illustration of the ALLOW / REVIEW / DENY idea behind the Sentinel authority gate. This runs only in your browser, using a few fixed keyword rules. It is not the Glass Box code, it calls no AI model, and nothing you type is sent anywhere.

Result appears here.

Interpretation is not authority, and model confidence does not become authority by itself.