I work on AI assurance and cloud security: measuring what a safety claim actually establishes, and where the evidence runs out.
Site → cubits11.github.io — technical claims there either link to public evidence or are marked as attested rather than shown.
-
CC-Framework — Python. If two AI content filters each fail on 10% of harmful inputs, multiplying gives 1% for the pair — but that assumes their failures are independent. The real rate lies somewhere in [0%, 10%], depending on how far the two failure sets overlap, and measuring each filter alone never tells you which. CC-Framework reports the range the measurements actually support instead of assuming independence. Worked example: When Marginals Are Not Enough.
-
Ghost-Ark — TypeScript, AWS CDK. At Penn State's S2 Lab. A verifier that tests how far a signed audit receipt actually identifies the execution it claims to describe, and where that identification breaks down.
-
Assay — private, early. What a verifier can and cannot conclude from AWS Nitro Enclave attestation.
Penn State CS '26, Cybersecurity minor · AWS Certified Cloud Practitioner and AI Practitioner · Philadelphia, PA.
Repositories marked Archived are earlier exploration kept as a record, not as current claims — some carry language I would not write today. Current work is what's pinned, plus the site above.