Eval Dashboard.

Fifteen questions run end-to-end through the same graph the chat page uses — nine answerable, two cross-company comparisons, two ambiguous, two unanswerable. Each is judged mechanically: gold-section citations, every figure re-verified, refusals where refusal is the only honest answer.

Measured on the bundled corpus (AAPL · MSFT · NVDA snapshot, Oct 2026). Session-ingested companies are not part of these measurements.