# Module 6 practice · account for selected evidence

Use [the lesson](../lessons/06-select-context-without-losing-evidence.md) before [the solutions](../solutions/06-context.md).

**A. Predict a budget.** Ranked passages A, B and C contain 150, 30 and 40 quote characters. The budget is 80. Under the taught whole-passage skip rule, which fit, how many quote characters are used, and which ID is skipped? If labeled required evidence is `{A,B}`, what is packet coverage? Can the selector legitimately read that relevant set to improve the reported comparison?

**B. Reproduce the loss.** Run `methods.py context`. Record selected IDs and labeled coverage at budgets 80, 120 and 400. Explain why the retrieval candidate set can remain identical. State three things excluded from the quote-character budget and what a later real-model input counter would need instead.

**C. Source ambiguity.** Two current public records share source `synthetic://hours` and versions `9` and `10`. They give different opening schedules. Explain the context rule's response and why sorting strings is not an authority policy. Name a different contradiction this rule would fail to detect.

**D. Duplicate text.** Two top-ranked IDs have exactly the same quote text, followed by a needed second-source passage. Explain what exact-text deduplication changes and what authority question it cannot settle. Would overlapping but non-identical chunks be removed by this implementation?

Self-review: predict the budget exactly, preserve denominators, explain authority limits and distinguish actual character selection from unrun neural token/position experiments.