Parse flow
Instant capture, quietly refined
A rule engine answers in milliseconds, and that answer is the product's answer. The on-device model refines in the background — and a small, honest indicator makes that improvement legible instead of silent.
The three moments
Capture
Instant result
The rule-engine parse appears the moment dictation ends. Save is live — nothing waits on the model.
~40 msRefine
“Refining…”
The model works in the background. A quiet pill says so, so the first pass reads as provisional — not as a poor capture.
~6 s · backgroundReveal
It gets better
Changed fields highlight briefly — you see the upgrade and credit the system. Your own edits are never overwritten.
perceptibleReview card · the same update, two beats
Left: right after dictation. The result is a first pass and Save is enabled — you are never blocked. Right: the refine has landed; only the fields the model improved are highlighted.
New update
Quick capture — refining“Vendor contract renewal, Sarah owns it, due end of August, waiting on legal redlines.”
Priority
Vendor contract renewal
Owner
Sarah
Due
Aug 31
Status
Waiting
New update
Updated“Vendor contract renewal, Sarah owns it, due end of August, waiting on legal redlines.”
Priority
Vendor contract renewal
Owner
Sarah
Due
Aug 31
Status
Blocked
Why it is held
Legal redlines outstanding
The model did not re-type the capture. It sharpened Waiting into Blocked and pulled the reason for the hold into its own field — the one you need when someone asks why.
Why it is built this way
The fast path is the real path. Most AI capture tools make you wait on the model, which turns a few-second habit into a several-second stare. TemporalNotes inverts it: the deterministic parse is the answer you can save immediately, and the model is an upgrade that arrives after you have already moved on.
The indicator is an honesty device. A first pass that silently changes under you feels like a bug. A first pass labeled refining is a promise kept a few seconds later — the same improvement, read as competence instead of instability.
Your edits win, always. If you correct a field before the refine lands, the model does not overwrite you. Refinement fills gaps and sharpens classifications; it never argues with the human.
None of it leaves the device. Both the rule engine and the model run locally. There is no API key, no round trip, and no transcript of your deals, personnel, or legal matters sitting on someone else's server.
The research behind the problem this solves: The Answerability Gap →