Capability elicitation

Testing prompts, examples, tools, sampling, or scaffolds to estimate the best performance a fixed model can produce on a task.

Observed capability depends on both the trained model and the method used to elicit a result. Worked examples, decomposition, tools, search, sampling, and scoring can reveal performance missed by one bare prompt. The procedure does not prove that a neat human-readable skill was already stored inside the weights; it establishes what the configured system can do under tested conditions.

Builder example

One failed prompt is weak evidence that a model cannot support a workflow. One successful scaffold is also incomplete evidence because it may be expensive, brittle, or tailored to the test. Compare realistic configurations and report both the best result and the resources required to obtain it.

Common confusion: Elicitation keeps model weights fixed. Fine-tuning changes them and is therefore a different intervention, even when both approaches improve observed performance.