Answers and explanations · Select context without losing evidence

A. A is skipped because 150 exceeds 80. B fits, leaving 50; C fits, leaving 10. The packet uses 70 quote characters. budget_skipped_ids contains A. Of required evidence {A,B}, only B survives, so packet coverage is 1/2. Using the relevance labels to select the answer would make the experiment depend on evaluation knowledge unavailable to the proposed runtime method. Labels belong in measurement, not this selector's inputs.

B. Budget80 selects none and has coverage0. Budget120 selects returns-v1 and has coverage0.5. Budget400 selects returns-v1 plus hours-v2 and has coverage1. The ranker runs before the budget, so these choices do not change its candidate list. JSON metadata/formatting, the user question and task instructions are excluded from quote-character accounting. A real model input counter would need the actual serialized request and actual tokenizer, plus output/interface reservations; no fixed ratio is asserted here.

C. Both versions are blocked from the selected context and the source/version conflict is recorded. The rule does not declare version10 authoritative. String order can even put 9 after 10, but numeric order would still not establish approval. Contradictory records under different source identifiers, or different content mistakenly sharing one version label, are not detected by this simple version-set rule. A review of the actual source policy is still needed.

D. Exact duplicates consume one context copy rather than two, potentially leaving room for the next source. The first ranked copy retains its provenance; the duplicate IDs are recorded. Identical wording alone does not identify the authoritative source, establish permission or justify merging two source histories. Non-identical overlapping chunks remain because the implemented comparison is exact string equality, not semantic deduplication.