Granite Bay High School · Granite Bay, California
I build reproducible systems that test what AI verification actually checked, then turn the result into evidence a stranger can inspect.
GitHub counts measured . Every other number links to inspectable work.
Selected work
Reproducible research
Executable research on verification boundaries: skipped coverage, self-selected trust anchors, policy scope, log integrity, and ambiguous evidence.
27 synthetic fixtures, six paired historical parser cases, and public CI.
Inspect the workWorking prototype
A Python prototype that separates evidence status from execution policy before a financial action can proceed.
PASS, FAIL, or INDETERMINATE evidence feeds a separate PERMIT, BLOCK, or ESCALATE decision.
Inspect the workDefect research
Minimal reproductions of evaluators and safety checks that can report success without completing the check their result appears to certify.
Each published case names its reproduction limits instead of generalizing from one defect.
Inspect the workNobulex architecture
The world can be ambiguous. The system’s response to that ambiguity does not have to be.
Bind the proposed action to the exact inputs, sources, timestamps, and identities it depends on.
Return PASS, FAIL, or INDETERMINATE. Missing evidence never becomes a clean result.
A bounded deterministic policy returns PERMIT, BLOCK, or ESCALATE.
Sign a receipt that binds the action, evidence references, policy version, and outcome.
External evidence
The public research distinguishes merged changes from endorsement and freshly reruns only the case it says it reruns.
The maintainer reproduced the failure, released a corrected verifier, and confirmed all four negative checks.
The fixture expectations, historical replay metadata, and deliberate-regression checks are inspectable.
Working standard
“If I cannot establish it, I should not return a clean result.”
I publish the command, expected result, and limitation beside each claim. Corrections stay visible.