Transparency

Transparency is not a disclosure policy — it is how the system works. Interface records the conditions under which every inference is produced because that record is what makes the inference useful beyond the moment of production. You can return to it, audit it, compare it against a different model or prompt version, hand it to a reviewer who wasn't present when it was made, or defend it to someone with legitimate grounds to question it. Without the record, none of that is possible.

Lipton's taxonomy of interpretability — simulatability, decomposability, algorithmic transparency — maps onto the different levels at which Interface provides inspection. Individual output level: what the model produced, over what source material, with what prompt. Model level: which model, which version, under what conditions. Process level: how the workflow was constructed and by whom, and which human review decisions were made along the way. GDPR Recital 71 establishes the legal floor for why this matters: individuals affected by automated decisions have a right to meaningful information about the logic involved. Interface is designed to meet that requirement structurally, not to treat it as a compliance minimum.

The practical consequence is that Interface records more than most tools, and makes more of that record surfaceable. Prompt versions are tracked. Framework versions are tracked. Model versions are tracked. Human evaluation decisions — judgments about whether an output is fit for purpose — are logged against the inference they relate to. The chain of evidence travels with the output rather than being discarded at the point of delivery.