The log

The Italian portfolio is not building three apps. It is building one known-word list per person, with evidence tiers and provenance: which words you can produce, which you only recognize, and which surface proved it. Every product accrues to that list and spends from it.

Saying it out loud exposed a gap. What ships today is one list per learner per surface, keyed to identifiers that never join. One person on three devices counts as three learners, Filotto builds its own list and never reads Paroletta’s, and the bridge between them was designed but not built. The number the portfolio grades itself on was summing anonymous fragments and calling that the asset.

So the list gets one home and one anchor. If the hypothesis under test is that the list is the valuable object, an architecture where the valuable object can be orphaned by a cleared cookie is not testing it.

Ten shipped things and three old names between them: Zen Space Labs (a defunct company, still the seller line on the App Store), Sierra Rails, and Paola Software. As of today there is one owner for all of it, called Enginery. It holds the iOS and Android apps, the web properties, the Work Context Protocol, and the client work.

The name is not decoration. Latin ingenium meant the cleverness begotten inside you, and it descends along two limbs: Italian kept ingegno, which still means both ingenuity of mind and a contrivance or machine, while English turned it into engine, kept the machine, and let the mind sense die out around 1600. Enginery is the English word left standing on that seam.

Products keep their own names. Enginery is the publisher, not a prefix.

Android was a generation behind: it matched the game as of Aug 12, not Aug 20. It was missing the farm card, both new modes, the per-artifact share codes, and the shared streak key. All of that was built and shipped in a single day, and the signed release went into Google Play closed-testing review under the Enginery developer account, with the store listing complete in two languages.

Two details that mattered more than they look. App signing enrolled by uploading the existing key rather than letting Google mint a new one, which would have orphaned every sideload install. And Play requires twelve testers to stay opted in for fourteen continuous days before an app can be promoted to production, so the real gate is recruiting, not code. Production lands around mid-September.

The plan said iOS would stay on TestFlight until Categorie and Vocabolario were finished, with one App Store submission in mid-September carrying everything. Instead the build went to review early with both modes aboard but switched off, which meant nothing about the shipped app changed except that it now contains two features nobody can see yet. Apple approved it today.

Two consequences worth naming. The App Store cohort clock starts now rather than in three weeks, so the activation and week-two-return numbers get an extra three weeks of real store traffic before the first grade. And the input-flow fix, which was one of the two known causes of people touching the game and never producing a word, is in front of store users now instead of in September.

Filotto keeps several tiers of evidence about a word. Some are strong: you answered it, or you marked it known on the calibration grid. One is a statistical inference from a recognition checklist, useful for placing a learner and weaker than it looks. The chapter generator had to say, in code, which tier the shipping gate measures against.

Both candidates were run against a volume already read and judged. The generous boundary scored it at 99.4 percent, comfortably easy. The narrow one scored it at 95.5 percent, and its reader had said that chapter read harder than usual. The generous gate would have called a chapter too easy on the evening its reader called it hard.

So coverage counts proved words only. A statistical inference about two thousand words is not two thousand words of comprehension.

The nightly reader is live: a chapter generated for one reader, verified against their known words, and mailed out. The question was how to tell whether anybody reads it.

A tracking pixel measures image loading, not reading. Apple Mail pre-fetches images for much of the plausible audience and other clients block them outright, so the number would be wrong in both directions at once. Instead every send mints a token, the email carries the title and a two-line lead-in, and the link goes to that reader’s own page. Fetching the page is the open, recorded server-side.

The cost is named in advance: the chapter text is not in the email, so a push surface became a pull surface. If opens come in under 50 percent, the first hypothesis is that this decision caused it, and the variant to test is putting the text back.