# Module 7 practice · claims and authored templates

Try [the lesson's](../lessons/07-draft-only-supported-claims.md) method before opening [the solutions](../solutions/07-generation.md).

**A. Observe three statuses.** Run `methods.py generation`. Explain why the reviewed fixture renders, an unknown question ID abstains and a missing source abstains. State which component wrote the returned sentence and whether an LLM ran.

**B. Repair a bundled claim.** A draft says: “Books can always be renewed for seven days and laptops can be borrowed for fourteen days.” Separate the factual points, compare them to the supplied sources and write a supported response without inventing a laptop policy. Keep the renewal condition attached to the permission.

**C. Spot a hidden classifier assumption.** A caller maps arbitrary natural-language questions to `q-renew` whenever they contain the word `days`. What goes wrong for the laptop question? Would matching the exact renewal source fingerprint prevent this wrong question mapping? Explain what the template guard does and does not establish.

**D. Design an optional model experiment on paper.** Specify input/output fields for a future model adapter, the version/configuration information to record and two answer failures to evaluate. Explain why a JSON schema and a “use only evidence” instruction cannot by themselves establish factual support. Do not obtain credentials, download a model or make a request.

Self-review: preserve the condition, avoid unsupported prohibitions, name the human-authored mechanism and keep the unimplemented classifier/model boundaries explicit.