Proof & Real Use
Everything on this tear sheet is read directly from database updates or from the latest Quality Lab run on disk. Anything not yet measured says so explicitly, rather than presenting a fabricated or estimated figure.
What it has actually produced
From the real dataset only. Demo sessions are excluded, so a visitor trying sample updates cannot move these figures.
Each one fanned out to several people in their preferred language
0 marked as sent
Distinct individuals across WhatsApp and email
Words delivered per word spoken, once an update exists
Minutes per update, before and after
The before figure is timed by hand on the old manual workflow. The after figure the app measures itself, from opening Speak to the final sent message.
Measured across updates written the manual way
Measured automatically per update
Requires both baseline and actual measurements
Requires both baseline and actual measurements
No time saving is shown because neither the baseline nor any elapsed time has been measured. Fabricating an estimate would compromise the integrity of every other claim on this page.
The latest Quality Lab run
A recorded suite of spoken updates run through the live pipeline, each graded against gold-standard recipients and extracted facts. Hard gates cannot be averaged out.
| Metric | Result | Target | Verdict |
|---|---|---|---|
| Invented factshard gate | 0 | 0 | met |
| Privacy heldhard gate | 100.0% | 100% | met |
| Must-know coverage | 42.7% | 95% | below target |
| Routing recall | 93.3% | 90% | met |
| Routing precision | 96.7% | 85% | met |
| Language and script | 100.0% | 100% | met |
| Channel fit | 100.0% | 95% | met |
3 of 20 cases passed every check.
How these numbers are produced
- Real-use totals are aggregates over the real dataset. No personal name, phone number, or email address is read or stored on this page.
- The accuracy table is the most recent Quality Lab run, which sends recorded updates through the identical pipeline used in production.
- Invented facts are counted by a deterministic verifier with no generative model in it. It maintains absolute authority over all draft releases.
- Anything not measured says “Not yet measured”. No figure on this tear sheet is estimated, rounded up, or carried over from unverified runs.