Harness
The software around an AI model. It decides what the model is given, what it may do, and which checks run on its work.
A harness is the software built around an AI model. It decides what material the model receives, what actions the product lets it take, and which checks run on its work. Two products can use the same model and behave differently because their harnesses differ.
Imagine a compliance officer testing two products that draft suitability reports. She gives each the same test file for Martin Doyle, with no attitude-to-risk questionnaire. The first is set to flag that missing document before it drafts. The other writes a report that treats Martin's risk level as settled.
The firm already uses templates, system access and file checks to control its work. A harness can put similar limits around what an AI product reads, writes and does. The comparison stops there: the firm may not be able to inspect every check in a product it buys. An update can change those checks without changing the model. Testing Martin's file shows whether the product flags this missing questionnaire, not which other errors it catches.
Not to be confused with
- The model produces the draft. The harness decides what the model receives and what the product does with its output. Model here never means a model portfolio.
- A guardrail is a limit or check within a harness. The harness can also supply context, instructions and connections to other systems.
- A vendor may call its harness a wrapper. That word has nothing to do with a tax wrapper such as an ISA or a SIPP.
Overheard
They both run on the same model. Why not take the cheaper one?
Same model, different harness. Only one flagged the missing questionnaire.
Where it sits
Harness
Copy this entry
Paste it into Word, Google Docs, Notion or Apple Notes. Headings and links come with it.
Sources
Your feedback
Could you explain harness to a colleague now?