One list per person, three doors into it

The 90-day OKRs were adopted on Aug 20. This is the state a day later, which is the start of the window rather than a grade. The first real grade is 2026-09-01, and most of what these objectives ask for is not readable before then.

Ahead of plan. Both new Paroletta modes are code-complete on all three platforms, ahead of the build weeks the calendar allowed for them. The iOS release was approved early, so the App Store cohort clock starts about three weeks sooner than planned and store users have the activation fix now instead of in September. Android went from indefinitely deferred to submitted for Play closed-testing review inside a single day, which added a whole platform the plan did not expect to have in this window. The measurement work that the objectives read from is shipped: the nightly graph snapshot, the visible farm on every platform, and per-artifact referral codes.

At risk, honestly. Two different metrics are both named “share arrivals” and disagree by a factor of five. Until that is reconciled, the distribution grade would be unreadable, so it is the first thing to fix before Sept 1. The nightly graph snapshot currently counts one product’s words and not Filotto’s, which understates the asset the first objective is about. The first live Salotto session is committed for late August and the invitation had not gone out as of this writing, which makes it the only commitment in the whole plan at risk purely from an unsent message. And the mode launches, the tester recruiting, and the first external channel post are all queued behind one Google review, which is nobody’s to hurry.

Targets get recalibrated on Sept 1, in writing, with the changes annotated. Not silently in between.

The deck makes one argument: build the asset, monetize later, and the asset is the per-learner known-word graph.

It starts from a problem the category has and nobody solves. Acquisition happens one step past what a specific learner already holds, not at the average level of a CEFR band. Measuring that properly is tedious enough that people quit, skipping it gives you population averages, and borrowing it means depending on a competitor. Meanwhile large language models made content free to regenerate, so the libraries incumbents spent a decade building are the part of the moat that is evaporating. What is left is the measurement, and everyone keeps theirs locked up.

The answer is to make the measurement the activity. Every product is something a learner would do anyway, built so that playing it emits vocabulary signal as a byproduct. Three doors, one graph underneath. The graph then buys what a fixed course cannot, starting with a chapter tonight that this reader can actually read.

The learner it serves is the one the category graduates and abandons: a few hundred words in, and nothing to read with them.

The deck itself is not posted here yet. It was built for a private room, and it gets a read-through before it gets a public URL.