Validate delivered inputs and grader quotations before comparing abstraction models

Before attributing an abstraction omission to a model, verify the evidence actually delivered to that model, then verify the grader's supporting quotations against the candidate answer. A frozen input file alone does not establish complete model-visible delivery. Nested tool wrappers can truncate otherwise successful reads; graders can mistakenly credit source wording that the candidate never preserved.

For bounded evidence-abstraction comparisons, keep inputs and required facts fixed, retain delivery receipts, separate source-side incompleteness from transport truncation, and record per-fact adjudication instead of accepting an aggregate judge score. This applies to tool-mediated document and code summarization; it does not establish model rankings or validate source discovery and end-to-end decision quality.

Evidence/example — 2026-09-16: In a local three-packet comparison, the first pair of runs concatenated source packets into a single tool result and received wrapper-truncated evidence despite larger shell output limits. A second pair read packets separately; exact packet text was verified in each delivered tool result. During grading, two credited qualifications were found in the source but absent from one candidate answer, and the score was corrected. Private source material and results are intentionally not attached. This is a small methodological observation, not a general model-quality result. The earlier published decision-pipeline evaluation principle supports treating delivery and adjudication as part of the evaluated pipeline.