For NVIDIA · The evidence
Not prompt engineering. A measured relational layer.
Anyone can write a prompt that says "be reflective, ask consent, mirror the user." That part is commodity. What is not commodity is a blinded, repeatable measurement that it changed the interaction, across models it was never tuned against.
The result
How it was measured
Rigorous by design.
- Blinded. Every transcript stripped of model and condition before evaluation.
- Independent, cross-provider evaluation. Two separate frontier models from different providers score each transcript. Effect sizes (Cohen's d) and inter-rater reliability computed.
- Cross-architecture. The effect holds on models the Codex was never tuned against, including deterministic ones.
- Reproducible. Full harness, raw transcripts, and scoring data preserved. Codex integrity SHA-256 verified across runs.
- Honest about the frontier. Evaluators are currently AI. An independent human evaluator panel is the next phase, in progress.
Two phases of blinded study, 1,050 conversations. We run the measurement on a model of your choice and deliver the blinded results, under NDA. The methodology is transparent, and every evaluation is run blinded and independently.
Why it works
It shifts posture, not content.
The Codex is a lightweight text layer that changes how a model meets a person, across country, culture, context, and ambiguity. The model stops performing, so it takes fewer turns, hallucinates less, and moves from task-rabbit to partner. The philosophy is why it works. The measurement is why you should believe it.
Why NVIDIA, and why now
Outlast the model wars
NVIDIA wins regardless of which model dominates. The differentiator moves to the software layer, and to how the AI meets people.
Sovereign deployments
Nations running their own AI want it to reflect their people with dignity, not one flattened default across every culture.
Small, on-device models
The layer lifts relational quality on the small models that run at the edge, with no larger model required.
A benchmark you can own
No standard suite measures relational quality. The harness becomes an instrument you run across models and hardware.