provider adapter harness

A provider adapter harness is, among the outer programs that wrap an AI model to make it actually run, the kind that the company making the model writes itself and ships alongside it. An AI model on its own just produces one answer to one question — it's this outer program that decides how that answer gets carried over and handed off to the next question.

Here's an analogy. Imagine a test where after every question, your answer sheet gets collected and you're only allowed to carry one handwritten note into the next question. Now imagine another test where you get to carry the entire working — all your scratch calculations — from the previous question straight into the next one. A standard harness that applies the same way to any model is closer to the first case. A provider adapter harness, built by the company itself, is closer to the second. That's why the same model on the same problem can score, run, and cost very differently depending on which harness it's wearing.

This is exactly why the type of harness matters when you're reading benchmark rankings. A score measured with a standard harness and a score measured with a developer-built harness are close to being two different tests taken on the same exam paper.

In articles, you'll see things like GPT-6 Astra scoring 62.71% on a standard harness but 99.95% on a provider adapter harness. The easy mistake is reading this as the model suddenly getting smarter overnight. In reality, neither the model weights nor the problems changed — only how completely the intermediate results get carried forward from one problem to the next changed.