You ask "what's tonight's hotel?" and the bot says the wrong place. Or "what's the load-in?" and it gives yesterday's time.
First — correct it
Text the agent the correct value directly:
- "The hotel tonight is the Marriott Downtown, not the Hilton"
- "Load-in is actually 2pm, not 1pm"
- "Tomorrow's van call is 9am, not 10:30"
The agent writes a user_correction row to day_facts. Future queries get the correct value. The correction beats any auto-extracted fact.
If the source is a Google Sheet, the agent ALSO writes the correction back to the source cell.
Then — figure out why it was wrong
A few possibilities:
1. The source document is wrong
Open the day sheet / advance / itinerary the bot extracted from. Is it actually right? If the doc says the Hilton but the band is at the Marriott, the doc is stale — get it updated.
2. The extractor mis-attributed
Sometimes the day_facts extractor stamps a fact on the wrong date. Example: a row in the itinerary lists "Drive to Amityville Music Hall, load-in 1pm" for an off-day, and the extractor pulls Amityville + 1pm load-in onto the off-day instead of attributing them to the next show.
The cleanup pattern:
- Text the agent the correction (becomes a user_correction row that wins forever)
- OR manually edit the source sheet to be less ambiguous
3. Multiple sources disagree
If you have a Master Tour day sheet AND a manual advance doc and they conflict, the extractor pulls both. The user_correction > manual > Master Tour sync > ingestion precedence resolves it (connect Master Tour to make it the trusted source), but the underlying ambiguity means a future ingestion could regress.
Best fix: pick one source as canonical, archive the other.
4. The query went to the wrong day
At 1 AM you ask "what's the hotel tonight?" — the agent applies the tour-day rule (before 4 AM = yesterday's tour day), so it gives YESTERDAY's hotel. Probably what you meant, but if you actually meant "tomorrow night's hotel", be explicit.
When in doubt, see what data the bot sees
Text: "what's the day_facts for today?" — the agent dumps its raw structured view of today, including every field that's set and its origin. You can spot which facts are wrong and which need correction.
Coverage page
/dashboard/coverage shows, for upcoming show days, which fields the agent has populated. Blank cells in the matrix flag gaps; oddly-populated cells (e.g. an off day with a venue) flag misattribution.