Open Decisions — Five Quick Calls + One Action

Decision 1 of 5

Fix the mispronounced words?

Sound & pronunciation data

🔊

What this is: Every word in the app has a "pronunciation guide" behind the scenes — it tells the read-aloud voice, and the on-screen pronunciation hint, exactly how to say the word. A check of every word containing the letter "e" found that 21.7% of them are wrong — the tool that first generated these guides never actually checked whether the vowel sound it picked was right. A batch of 10 corrected words has already been double-checked and is sitting ready to load into the database as one clean update — nothing has been changed yet.

Word	Says now	Corrected to	What changed
cepat	/ˈtʃe.pat/	/tʃə.ˈpat/	e → schwa, stress shifts
dekat	/ˈde.kat/	/də.ˈkat/	e → schwa, stress shifts
penting	/ˈpen.tiŋ/	/pən.ˈtiŋ/	e → schwa, stress shifts
karena	/ka.ˈre.na/	/ka.ˈrə.na/	e → schwa, stress shifts
lebih	/ˈle.bih/	/lə.ˈbih/	e → schwa, stress shifts
rezeki	/rə.ˈze.ki/	/rə.ˈzə.ki/	e → schwa, stress shifts
terus	/ˈte.rus/	/tə.ˈrus/	e → schwa, stress shifts
berhenti	/bər.ˈhen.ti/	/bər.hən.ˈti/	e → schwa, stress shifts
memang	/mə.ˈmaŋ/	/ˈme.maŋ/	reverse case — schwa → e (new finding, not yet ear-checked)
mereka	/mə.rə.ˈka/	/mə.ˈre.ka/	reverse case — schwa → e (new finding, not yet ear-checked)

A — apply this 10-word batch to the database now.
B — hold off, don't apply yet.

👂

Three words are being left alone — your call if that's still right. These three showed the exact same kind of vowel mismatch, but you already ear-approved keeping them as they are on 2026-07-17. They are NOT included in the apply/hold decision above — this is just a chance to flag if your ear has changed on any of them.

kena — app says /ˈke.na/, reference says /kəˈna/ — kept as-is
empat — app says /ˈem.pat/, reference says /əmˈpat/ — kept as-is
sedih — app says /ˈse.dih/, reference says /səˈdih/ — kept as-is

Decision 2 of 5

Which timer math drives the "~X min left" countdown?

Study-session pacing

⏱️

What this is: While you're studying, the app shows a small pill estimating how many minutes are left in the session. To do that, it needs one number: "how long does answering one card typically take." Two candidate numbers were measured from real usage: the typical (median) answer time — 8.9 seconds — versus a trimmed average (extreme outliers dropped, then averaged) — 16 seconds. Whichever one you pick becomes the number the countdown is built on.

A — use the median, 8.9 seconds. Resists getting thrown off by a few very slow or very fast cards.
B — use the trimmed mean, 16 seconds.

AI recommendation: A. The median naturally resists outliers, which is exactly the failure mode a pacing estimate needs to avoid.

Decision 3 of 5

Approve the standing testing checklist?

QA process

✅

What this is: There's a live page that tracks what to try on the app right now, what you've already confirmed works, what's a known issue that doesn't need re-reporting, and what's been asked for but isn't built yet. It keeps updating as things ship — approving it means this becomes the standing process going forward, rather than a fresh checkpoint every time something needs testing.

Test now (6)Game end-screens consistent · no mid-game scores/streaks · Context Fill chip cleanup · card reveal info trimmed · My Cards category filter · richer dictionary content

Confirmed (1)Profile switcher — no more hydration errors

Known issue (1)Home screen tile borders look inconsistent — already being investigated, no need to re-report

Coming soon (3)Estimated-time pill on the Carousel (this is Decision 2 above) · Reveal button restyle · Profile switcher "+ New" repositioned

Open the full testing log →

A — approve this as the standing checklist.
B — changes needed (say what in the note).

Decision 4 of 5

Where do the new process rules get written down?

Fleet process — how status gets reported to you

🧩

What this is: A recent audit found a real decision of yours got left off a status update. Two process fixes and one communication rule came out of it, and all three need a permanent home so they actually get followed next time:

Rule F-A — any document that contains open decisions for you must also produce a checkpoint page like this one, in the same work session it's written. No more decisions sitting silently in a doc you never see.
Rule F-B — whenever the fleet answers "what's waiting on you," it must be freshly recalculated from the fleet's running notes file every time — never answered from memory. Memory is how the decision got left off last time.
Communication rule — your own instruction: "be extremely clear." Short, plain sentences, one idea per sentence.

A — write these into the fleet's own lessons log + its operating manual. Recommended — these are rules about how the fleet runs itself, not about you personally.
B — write these into your global instructions file instead.
C — write them in both places.

AI recommendation: A. These are fleet-process rules — they belong where the fleet's own operating rules already live.

Decision 5 of 5

Confirm how word-forms relate to their meanings

Content model — forms vs. senses

🔗

What this is: You said the data for a word-form (like menurunkan, a built form of the root turun) is "different and the same" as the meaning it comes from. Here's that restated as four plain rules — confirm this is what you meant:

1. A form (like menurunkan from turun) gets its meaning from one parent sense — that is the "same" part.
2. The form is also its own item — its own examples, its own properties — that is the "different" part.
3. Every piece of forms enrichment must validate twice: against the form itself, and against the parent sense.
4. A form without a parent-sense link is invalid by construction.

A — confirmed as stated.
B — needs correction (say what in the note).

Action needed

Rotate the leaked ElevenLabs key

Security — only you can do this step

🔑

What happened: The API key for the text-to-speech (read-aloud voice) service got accidentally printed into a work log yesterday. Once a secret like that has been exposed, replacing it is the only real fix — deleting the stray copy alone isn't enough.

What to do: Generate a new key in the ElevenLabs dashboard (elevenlabs.io), then it needs to be saved into the fleet's password vault under the name 11-labs-eleven-labs-free-account-TTS-Soundeffects. This has to be you — no one else has dashboard access.

Five small decisions and one action