Wading has one number that matters: the share of your own conversation that you produced in the language you're learning. Not words seen, not lessons finished, not days in a row. What you said.
Everything else in the app is instrumentation around that number. So it's worth being exact about what does and doesn't count toward it — and about a case we got wrong.
The bug
Underneath the composer there are quick replies: two or three suggested things you might say next, in the language you're learning. They exist because starting a sentence in a new language is the hard part, and a nervous learner with nothing to say tends to close the app rather than struggle. Tapping one keeps the conversation moving.
Those taps were counting as production.
You'd tap a chip reading hace calor and the system would record that you produced two words of Spanish. Which is nonsense. You recognised a phrase and judged it apt — real, but a completely different skill from reaching for a word and finding it there. The words were ours. We wrote them, put them on a button, and then congratulated you for pressing it.
Nobody reported this as a bug, because it doesn't look like one. It looks like progress.
The fix, and what it costs
A tapped suggestion now records no production at all — and specifically not a measured zero. Those are different claims. "You produced no target language this turn" is a data point about you. "Nothing measurable happened here" is a statement about the turn. Conflating them would drag averages down as dishonestly as the old behaviour pushed them up.
The word-level signal still counts, capped. Reading a chip and choosing it correctly is evidence you understood it, so words can still advance from met to recognized to understood off a tap. They can't reach producing. That stage means you said it yourself.
The visible consequence: if you lean on the suggestions, your percentage is now lower than it was last week. Nothing about your Spanish changed. The number just stopped flattering you.
Why we did it anyway
Partly because a progress metric you can game by tapping is a progress metric that will, eventually, teach you to tap. Any measure that's easier to satisfy than the thing it measures ends up replacing it. Adding "press the button, watch the number rise" would have made the app worse in the specific way we built it to avoid.
And partly for a less high-minded reason: we're talking to a university about running an efficacy study, and this measure is the one genuinely novel thing we'd be offering — continuous, objective, token-level data on what learners actually produce, across a whole term, rather than a handful of proctored assessments. That's only worth anything if the number means what it says. A metric quietly inflated by UI taps isn't a research contribution, it's a mistake with a graph attached.
Better to find that ourselves than to have someone find it for us.
What still counts
Everything you typed or said, whether or not it was correct. Production is measured on what you produced, not on getting it right — a mangled sentence you reached for is worth more than a perfect one you tapped. Corrections arrive warmly and separately, and they never dock you.
Spoken turns count the same way, from what the transcript heard you say.
And the claims are audited. The model proposes which of your words were target-language; the server checks each claim against your actual message as whole-token sequences, so casa can't quietly verify itself against the English "castle." A wrong keep here inflates a stored number that can never be repaired later, which is exactly the class of error we'd just shipped.
The suggestions aren't going anywhere. Keeping a hesitant learner talking is worth far more than the credit it no longer earns.