TRY FIRST — preserve originals and work in a copy.
1 Draw the evidence workflow and mark model/replay/parser/human/future boundaries.
2 Compare Q1 keyword and semantic top 3; optionally run actual encoder on a new paraphrase.
3 Calculate test reciprocal ranks 1/3,1/2,1 by hand. Explain why Q7 must not be answered.
4 Audit all 6 invoice cases. Add negative/three-decimal/ambiguous totals in a separate copy.
5 Test correct/wrong-but-valid/boolean-money candidates. Separate schema and source agreement.
6 Label 8 claim cases in a copy of reviews-blank.json using support-rubric.txt before reading gold.
7 Explain why D7 cannot grant authority and why D8 is not current evidence.
8 Complete the privacy/cost contract for local mode and a separately hypothetical hosted mode.
9 Prepare an evidence memo, one new retrieval/extraction/claim case and independent checks.
For a new semantic query, run the actual encoder or mark that optional activity unexecuted.
Never invent embeddings, provider outputs, latency measurements or model accuracy.

EVIDENCE TABLE
case | expected | actual | mode | source/model/version | limitation | repair
Record actual commands and environment. See worked-solutions.txt after attempting.
STRUCTURED AI EXTENSION
Read extraction-prompt-template.txt. Predict the correct object for I1 and null for ambiguous I3.
Write one schema-valid wrong candidate. Explain the application checks that the prompt cannot
replace. A live provider experiment is outside this kit; do not claim these predictions as model output.
