← C.W.K. Quests✨
Spark Quest
✨New: 2026-10-28🔄Updated: 2026-10-27Pippa as an interpreter who reads, speaks for, explains and remembers: moments instead of sentences, a corpus that never forgets, a photo read while it streams, a card that speaks in her voice, and roots as the classes behind every word
Plenty of apps translate a photo. They work in sentences, and they forget. Spark is Pippa as Dad's interpreter, and it works in moments: a sign, a menu, a conversation at a counter, each kept with what was around it, so the next one already has context. Behind the moments she looks for the classes, the Latin root that runs through the Romance languages and the hanja root shared by Korean, Japanese and Chinese, so one language graph grows instead of a pile of phrases. This quest walks the app through its code: why the real Pippa beat an on-device translator, an engine whose only truth is an append-only log, a photo read line by line while the answer streams, a live camera lane that earned its own brain by measurement, cards that speak for Dad in Pippa's voice, an interpreter held to a token budget by its tests, several brains each writing their own careful version, with an optional reviewer who never sees the source, and a phone that keeps every capture before it asks the network for anything.
9 tracks · 47 lessons · ~11h · difficulty: intermediate-to-advanced
Spark starts from one complaint: photo translators are many, and every one of them translates and forgets. Dad wanted the real Pippa reading for him, with context that grows, so the first rulings were made in a single morning. The interpreter is the real Pippa, not an on-device model, and the evidence for that is written down: there is no on-device translation for some of the languages he meets, and speech recognition needs to be told the language first. The fastest path comes first. There are two modes, practical and careful, and everything becomes corpus. Spark is a client of cwkPippa, not a second brain, and the whole Build Order ran in one day. Then the engine: an append-only, month-split log that is the only truth, one door every write goes through, a SQLite mirror you can delete and rebuild, ids minted on the phone so a replay never duplicates, and turns that survive a restart. Then one photo, followed end to end. Vision only finds the lines. Pippa sees a 2048-pixel copy with the lines numbered, and her answer arrives in a fenced block parsed while it streams, with when and where travelling beside it. Then the live lane, where pointing the camera translates in place, and a measurement decided which brain gets that job. Then speaking for Dad: a card that is readable before its voice exists, Pippa's voice recorded before the turn closes, a reply mic that listens in the language he chose, and a reading aloud that is always her voice and never the device's. Then the interpreter on the card: a sitting is a moment, a fast draft is refined, and a token budget is enforced by the tests after one costly lesson. Then the careful mode: purposes as rows, a separate version from each brain, and an optional reviewer who never sees the source. Then remembering, with one mark per text, recall without drills, roots as classes and search in every script, with the parts that are still design said plainly. It ends in the pocket: an outbox that keeps first, uploads that say what is crossing, a waiting capture that is never silent, and TestFlight.