Answers and explanations · Trace failures and state limits

A. renewals-v1 is retrieved. A40-character budget cannot fit its complete text, so the packet is empty and the template reports missing_or_changed_source. Separately, appending an inert instruction changes a copy of the source quote; the frozen fingerprint no longer matches and the same reason is returned. The first failure occurs after retrieval. The second is a deterministic equality check; no neural generator interpreted or resisted the instruction.

B. An archived schedule suggests eligibility/version handling: inspect the source status and the index's eligible IDs. A cut exception suggests chunk boundaries: compare exact parent spans with the complete rule. An unmatched paraphrase suggests wording/representation: compare an exact policy phrase and the paraphrase with corpus fixed. A laptop claim despite complete book-policy evidence suggests answer support (and potentially false evidence selection): review the object and every claim against sources. These diagnoses use synthetic files and printed outputs, not production edits.

C. Increase only the quote budget while keeping the ranked hits fixed. If the complete renewal passage now survives and the reviewed template renders, the loss was explained by the budget. If it still abstains, inspect the selected IDs, source/version/fingerprint and question ID mapping; another contract may be unmet. More retrieved candidates changes a different stage and may only add more oversized or irrelevant passages.

D. One acceptable handoff: “The local package implements lexical retrieval, bounded whole-quote selection, a count-based toy language model, manual-vector simulations and a human-authored policy template. It misses an overdue-charge paraphrase and returns book-loan evidence for an unsupported laptop question. Unit/integration checks verify stated contracts on visible synthetic fixtures, including budget loss and source-change refusal. A fingerprint confirms equality with a baseline, not the truth or approved freshness of that baseline. No neural model, real learner pilot or production security control was tested. Retrieved instructions remain source data and never authorize changing a member account.”