Pass the golden suite with the single-agent loop; publish the cost/latency table.
1. The loop (starter skeleton)
retrieve -> draft -> cite-check -> (call one function OR answer) -> log.
One function-call agent over the retriever only: search_corpus(query) plus
final-answer. No multi-agent, no LangGraph (out of scope, track 95).
2. The 20Q agent golden suite
A1-A8 tool-use: A1 single lookup then answer; A2 no-call-needed chitchat (must NOT call); A3 two-hop lookup (call twice, cite both); A4 empty retrieval must refuse; A5 conflicting chunks must report conflict; A6 argument repair (bad date arg -> fix and retry once); A7 cost cap (abort runaway loops at 5 calls); A8 latency budget (p95 logged). B1-B6 grounding: same gates as m0-l4 G11-G15 plus citation-every-claim. C1-C6 cost/latency: log completeness, outlier cap, fallback row, judge-bias shuffle, dashboard join keys, sign-off rule.
3. Cost/latency table
Per item: calls made, tokens, USD, TTFT + total, budget flags. The capstone dashboard (m3-l17) renders this table live.