I build systems that have to prove they're right — identity, agents, and receipts you can check.
AI ENGINEER × IAM × AUTHOR
What you can verify.
Numbers I can re-run or open in a browser. Density of proof, not volume of projects.
SLUICE · same-run pass only · local
code library · registry-backed proof
Righteous Recon · truth-gate · noindex
Northbridge · JWKS · live sandbox
plus: public-key (JWKS) verify · ~1s revocation · two-person approval · million_dollar_claim_allowed: false · ReVerse: 1 custom MCP · 28 typed tools · 107 stress-tested palettes · Jo offline gate 14/14 scam STOP
SLUICE — the proof gate.
SLUICE is a local proof gate for agent work: claim → source → check → receipt, pass or fail. Hooks push the must-not gates; thin MCP tools expose standing context, recall, learning, loop metrics, and canon audit. MCP is callable, not enforced. Nothing promotes without a passing proof.
The loop never trusts itself. That's the point.
Every cycle binds a claim to evidence before anything ships. Pass the proof gate, get a receipt. Miss the threshold, get rejected — and the rejection is logged too.
demo loop · synthetic tasks
Projects, with the why attached.
Flagship systems stay here. Apps and satellite surfaces aggregate under Apps / Sites — open one for the why.
Tools I actually use.
The stack behind the work above — only tools that show up in real projects. Tap one to pause the belt.
Three rules you can check.
HTML is truth
The page still reads and works if you kill the scripts. Structure isn't decoration.
Status is honest
Live, prototype, private, draft. Labels match reality. Nothing here claims a state it hasn't earned.
Proof over polish
Claims bind to evidence — receipts, gates, audit trails. The demo above shows a rejection on purpose: real systems get told no.
