In order of effect we do three things:
1. Field-aware extraction, not transcribe-then-parse. the model knows
Before field X expects a datetime, field Y expects a phone,
Field Z is open-ended. So when someone says “Yeah Tuesday-ish,
Not really, Wednesday morning works better” – the sign is anchored
“What is the last desired date time?” No “What did this person do?”
Say?” Filler words and false starts are filtered out as noise because
They do not match the schema of the field.
2. The last statement wins for contradictions. if someone changes
direction in the middle of the sentence, we bias toward the most recent declarative
claim. “Email is Maya on Gmail – no wait, Maya on ACME.com” →
maya@acme.com. This also matches the way humans hear.
3. Trust-gaining confirmation phase. Each extracted field comes
Back with a trust value. Above ~0.85, it automatically fills and
The respondent sees it already filled in (still editable). Below that, we
Show “We think you meant X – is that correct?” never less confident
Silently writes incorrect data; It asks.
<a href