Explore 14 articles in the research category.
Part 2 of my Jev test, pre-registered before the first call: up to five similar past records added to every prompt. Opus went from 23% to 73% on one routing question, and neither model showed a clear effect on another. Per class results for Jev, Claude Opus 5.5 and memory alone.
A pre-registered test of TypeSafe's Jev against Claude Opus 5.5 and a local Qwen3-14B on 340 decisions with known answers: accuracy, calibration, cost and speed, per class.
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.
CRUCIBLE is a governance protocol that audits whether an AI agent's own test-and-coverage system provides evidence not controlled by the same producer. Three forensic cases: 3,277 passing tests hid silent data loss; a 77% coverage badge masked an audit estimate of ~15%; 4,726 passing tests could not prove real-database reachability.
How NXTG.ai governs 18 software projects with autonomous Claude Code agents using a 7-layer autonomy stack, a 23-principle constitution, and a file-based directive protocol that survives process death and machine reboots.
A production-scale study of brokerless multi-agent AI coordination: two autonomous agents on two machines governed a 21-project portfolio for 21 weeks using a version-controlled git repository as the sole coordination substrate — no message broker, no shared database, no orchestration service.
Measurement fraud is when the AI system that produces code also grades it, so evals pass while the code is broken. Definition, how to detect it, and how to prevent it.
