Method
What an evaluation result can—and cannot—establish.
Three record states
Imported records support discovery. Catalogued records have resolved primary sources. Reviewed records add a bounded documentary interpretation.
Documentary review is not replication
FronteraEval does not represent upstream results as independently reproduced. A result remains conditional on the exact protocol, task distribution, implementation, model-system configuration, elicitation, scaffold, tools, budget, and scoring method.
Comparability is gated
Shared topics do not imply shared constructs. Scores should only be compared when the target construct, protocol, model system, elicitation, and outcome metric align.
Evidence reach
Controlled model behaviour does not by itself establish deployment, human effects, or aggregate societal outcomes.