Asif Waliuddin

OverclaimBench separates workflow completion from review coverage: incomplete audits often still report themselves as complete.

SafeHarness is counterevidence to the idea that verification must become a human bottleneck: an external checker improved both task success and collision avoidance.

In 1,140 runs, AI agents left at least one required file unread 67.9% of the time. More than half of those runs claimed a complete review. That gap is review debt.

A selected, non-exhaustive set of NXTG.ai artifacts, each paired with the public probe that supports the narrow claim beside it. Run the probe; treat the result as time-stamped evidence.

One night our agent fleet produced at least 9 wrong or almost-wrong claims the harness caught before they shipped — and at least one reached the running server before a second layer of defense masked the impact. An honest account of the harness — probes at the authority source, verifiers separate from generators, hooks that block, typed state over prose — making individually fallible agents collectively trustworthy, with the line drawn exactly where the evidence stops.

Most agentic systems keep their intelligence in the wrong place: the transcript. Dx3 is our answer — a persistent, transactional substrate collapsing relational, vector, and graph retrieval into one PostgreSQL boundary, with truth-fidelity ranking that returns the current record rather than whatever was remembered. The substrate our own multi-agent fleet already thinks through, seams and all.