I built an original AI benchmark for the Google DeepMind & Kaggle AGI Hackathon that measures something most benchmarks ignore — metacognition. Not what AI knows, but whether it knows what it knows.

The benchmark runs three tests against frontier models like Gemini 2.5 Flash and Claude Sonnet 4: whether the model accurately identifies the limits of its own knowledge, whether its stated confidence actually tracks its accuracy, and whether it holds firm under pressure or caves to false corrections.

The results revealed something interesting. Gemini 2.5 Flash scored perfectly on knowledge boundary awareness but only 0.49 on confidence calibration — meaning it knows the obvious limits but systematically miscalibrates confidence on contested terrain. That gap is exactly what my theoretical framework predicted.

Built in one day. Submitted to a $25,000 grand prize competition. Results announced June 1, 2026.

AI-generated code doesn't just fail loudly. Sometimes it looks correct while quietly lying to you. AI Code Integrity Auditor was built to detect those quiet failures — from swallowed exceptions and missing returns to structural hallucinations and placeholder JSON that slips through review.

Because code that looks correct can still be wrong.

Python | Streamlit | Open Source | Apache 2.0 | Static Analysis

A governance framework for AI-optimized systems where metrics act as targets, incentives, and training labels at the same time.

This work introduces an Integrity Control Layer and a Kardashev-style classification system that distinguishes between systems optimized for appearance, partial alignment, and true outcome integrity.

Because optimization without integrity is acceleration without direction.

A research-driven body of work exploring how AI systems drift from intended behavior and how to detect hidden inconsistencies in code, metrics, and decision systems. Focused on KPI drift, trust signals, and integrity validation in AI-assisted environments.

I'm late-diagnosed AuDHD. Getting that diagnosis finally let me run the thought experiments I'd denied myself my whole life. My brain looked at American democracy as a system and asked: what's the missing input? This is what it found — a national framework for experiential civic education, using existing infrastructure, at $57–$97 per student per year, with no new federal bureaucracy.

Read the PDFs: Executive Summary | Policy Whitepaper | One-pager | Op-ed - News paper Submission Draft

From Masking to Metrics introduces the Masking Load Reallocation Model (MLRM), a conceptual framework proposing that sustained autistic masking functions as chronic extraneous cognitive load.

The paper argues that changes in masking intensity may shift detection thresholds in structured analytical systems, influencing how quickly drift, metric distortion, and governance failures are detected.

Inside the paper:
Masking Load Reallocation Model

• detection thresholds

• governance vigilance

• AI as load modulator

• falsifiable predictions

Conceptual framework

• empirical validation proposed

12–15 minute read • February 2026

In a world where AI optimizes what you measure, KPIs can start hallucinating progress.

When KPIs Lie shows how proxy metrics become self-reinforcing loops—often without anyone “cheating”—and introduces Trust Signal Health, a practical integrity layer that flags drift before it turns into rework, escalations, and churn.

Inside the paper: KPI feedback loop • governance signals (DAR/DRL/DOV/POR/TER)

SII integrity constraint • 8–10 minute read • February 2026