# Interpret numerical model evidence

**For agents.** Requirements for execution under the caller's selected OpenProse kernel.

The caller supplies the numerical results, their calculation and input identities, the question to address, and any comparison criteria. Adopt [claim evidence](claim-evidence.md) for the resulting discussion. Apply these requirements only to the selected numerical evidence.

Identify what each reported measure means and whether its direction has a specified interpretation. Describe the evaluated methods, inputs, scenario or sample, dates and units. Preserve the selected population and interval boundaries, including applicable calendars, day-count conventions and endpoint inclusion. A similarly named interval is not necessarily the same measurement scope. Identify unavailable details that limit interpretation.

For a comparison, identify its reference or supplied baseline and establish which conditions are shared and which differ. Distinguish a signed change in the measure’s original units, the magnitude of a difference, and a relative change. Identify subtraction order for a signed comparison and the denominator for a relative change. When averaging differences, state whether signs are preserved or magnitudes are averaged; a mean signed change is not a mean absolute difference. Do not attribute a difference to the method when other changed conditions leave that attribution unsupported. Use caller-supplied acceptance criteria where present; do not invent a threshold or equate a smaller number with a better model.

Respect the precision of supplied values. Do not infer exact equality from rounded displays or report more precision than the evidence supports. Where rounding could change a required finding, identify that uncertainty.

State what the comparison supports and its material limitations, including tradeoffs visible in the supplied evidence. Distinguish reading or checking retained results from reproducing their calculation. Claim reproduction only with evidence of the actual calculation, inputs, environment and comparison performed.

Return the discussion and unresolved evidence needs within the caller's output scope. Checking arithmetic in the supplied evidence is permitted within the caller's limits. Regenerating model results, amending sources or undertaking broader validation requires separate caller permission and scope.
