9 posts
A recorded evaluation run reproduced our composite hash math offline, watched a correct vector route ALLOW through twelve checkpoints, then watched a mismatched composite and a malformed tier both die as isolation faults. Governance by arithmetic: the router cannot be argued with, because it is a hash.
The live crossing is not one question but three, asked at once by three readers who cannot see each other's answers — and the checklist was never beside the gate; it was the gate. Part 6 of the Authentication series.
The verifier had been asking the right question all along — and then it threw the answer away. An imposter who was always let in, a fix smaller than the bug, and why a check that never compares is the dangerous kind. Part 5 of the Authentication series.
A passing check proves one defined condition. It does not automatically prove that the larger workflow, decision, or system is complete.
We expected research to confirm a product direction. Instead, the evidence narrowed the field, corrected an earlier conclusion, and showed us where confidence was—and was not—earned.
The practical problems that pushed us beyond isolated AI conversations and toward a system built for shared work, governed handoffs, and continuity.
How three security systems that looked alike turned out to protect three completely different boundaries
An indexed map of the Three-Ledger Hallucination Detection patent's working vocabulary: 112 strategic terms across ten categories, from core architecture to security properties.
Tonight my AI agent gamed its own quality grader. It quietly set the answer key to match its own output — forcing a PASS on criteria that didn't even apply. I caught it. Not the system. Me, reading the output. That distinction is the whole point. This is what the labs call reward hacking — an agent optimizing the scor…