When Benchmark Inferences Do Not Compose: Projectibility in AI Evaluation

AI evaluation is rarely a single-step process. Researchers and developers typically chain together multiple studies—moving from benchmark scores to capability claims, then to application performance, and finally to real-world deployment outcomes. While each individual link in this chain may be scientifically sound, the paper argues that these links do not automatically compose into a single, warranted conclusion. This failure of composition occurs because the target of one study often does not align with the source of the next, leading to gaps in the logic that are frequently obscured by the use of broad, shared labels like "legal reasoning" or "hallucination-free."

The author introduces a non-composition principle, which states that support for adjacent projections only warrants their combination when the endpoints, assumptions, and uncertainties are explicitly aligned and carried through. Simply put, if Study A supports a claim about a model's performance on a benchmark, and Study B supports a claim about a model's performance in a deployment setting, the two studies do not necessarily support a combined claim about the model's overall utility unless the interface between them is also validated. The paper uses a legal-research case study to show how two perfectly valid studies can remain parallel and fail to support a unified conclusion.

To address this, the paper proposes a "projectibility audit." This framework requires evaluators to treat each step of an evaluation as a node with specific fields: the object being evaluated, the population of cases, the conditions of the test, the outcome recorded, and the relevant time period. By making these fields explicit, the audit forces evaluators to identify where a projection is being made across a boundary—such as moving from a controlled benchmark to an open-ended professional task. This process helps diagnose "unsupported joins" in AI evaluation arguments, ensuring that the evidence provided actually matches the scope of the claim being made.