Choice copy was comparing the search window, so every answer said the range had not changed. Block-scan summed raw scores, so longer afternoon windows won before any evidence difference.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep askable cards after exhaustion, explain each probe, read the adopted credible range in reports and chat, and compare declared periods before the minute grid when the clock is unknown.
Co-authored-by: Cursor <cursoragent@cursor.com>
Implement offset and softmax relative-support scales and a public holdout calibration script, but leave the default proportional because no scheme passed the coverage-and-width gate.
Co-authored-by: Cursor <cursoragent@cursor.com>
The first-round title guard compared the pre-RPC snapshot, so append_consultation_question had already rewritten the title and the model name never landed.
Co-authored-by: Cursor <cursoragent@cursor.com>
The four account panes now share a fixed frame, chart profiles open as list then detail, and membership lives in the homepage dialog instead of a separate page.
Co-authored-by: Cursor <cursoragent@cursor.com>
Session history was silently clipped to the first 4000 characters of the last 12 messages, so follow-ups could not see timing or audit tables. Keep an append-only tail plus a checkpoint summary, retry overflow in the same request, and expose cache hit rate in admin usage.
Co-authored-by: Cursor <cursoragent@cursor.com>
Enter now queues one follow-up instead of dropping it, and a rectification stop leaves the streamed reply with a grey notice instead of an error alert.
Co-authored-by: Cursor <cursoragent@cursor.com>
Separated candidates were treated as adoptable while finance, relocation, health, or occupation collect was still pending, so clickable cards and a verification report appeared before those questions were asked.
Co-authored-by: Cursor <cursoragent@cursor.com>
A choice-card yes/weak_yes already told us the domain happened, so skip the same-domain "which year" collect without writing the ledger.
Co-authored-by: Cursor <cursoragent@cursor.com>
Family coverage still continues dated collection when nothing can split leftover minutes, but an already-built choice card must not be displaced by spoken finance collect.
Co-authored-by: Cursor <cursoragent@cursor.com>
Family "没有" was treated as adopt-ready once the training gate was open. Remaining dated domains still distinguish leftover minutes, so persist those questions instead of writing the adopt narration.
Co-authored-by: Cursor <cursoragent@cursor.com>
The time card stayed under an earlier collect question, and the close said to just use a time for now. Keep the card on the delivery turn and present the analysis as this round's result.
Co-authored-by: Cursor <cursoragent@cursor.com>
Adopted reverse-verify repeated already-asked collect stems, treated 「这题跳过」 as stopping the case, and promised holdout/OOS checks that never ran. Collect spoken stems no longer prefix a year.
Co-authored-by: Cursor <cursoragent@cursor.com>
The early-exit reused the pre-click dossier, so the agent never ran on the common path. Use this-turn decision, log agent vs template outcome, and time out after 8s.
Co-authored-by: Cursor <cursoragent@cursor.com>
When leftover probes cannot split adjacent minutes, skip frameless follow-ups and let a no-tool agent explain the stop. Distinguish-card "no" no longer closes a whole evidence domain.
Co-authored-by: Cursor <cursoragent@cursor.com>
Jesse_Chen