Twelve months ago, on the morning the engine first ran end to end, we wrote a short note saying it was not yet good. We meant it. The note ended with a forecast that v1 would ship in Q1 and a promise to keep the alpha-era voice on every post until the work had earned a different one. This is a year-end note on the same terms.
Year One is over this week, by the calendar we have been keeping. The first paying candidate sat down in front of the engine on March 4. Today there are roughly twenty-one hundred. The first Mock Exam attempt was completed on September 25 — three days after public beta opened — and by Sunday night the total had crossed thirteen hundred, which is the threshold we said we would wait for before promoting the surface out of beta. That promotion ships in January.
What shipped, in order
Each of these was its own note in the newsroom. The list, for the record:
February. The diagnostic, free, 25 minutes, returning a banded score and a single sub-skill to work on first. February. The retirement of the legacy GMAT closes the transition window. March. Engine v1 — proper 3PL Item Response Theory, Fisher- information selection, a 2,400-item pool. Pro 1-month ships at $199 a month against it. April. Pacing as a first-class signal inside the loop; the back- third accuracy gap closes by nine points. May. The calibration pipeline moves from manual to nightly; usable pool grows by 19%. June. Pro 4-month at $599 for ninety days. July. The "Focus Edition" label drops from GMAC's materials; the test is just "the GMAT" from then on. August. GMAC's percentile table refresh compresses the top of the curve; the engine's percentile column updates against the new table. September. The Mock Exam in public beta. October. Pro 6-month at $699 for half a year. November. The first written guarantee — diagnostic-banded lifts, three eligibility checks, extended access as remedy.
What we got wrong
Three things, briefly. Year-end notes that only celebrate the work that landed are the worst category of corporate document. We will keep the receipts on the work that did not.
We held the engine in alpha three months longer than we should have. We told ourselves the calibration pool was not wide enough, and we kept saying it through January and February while we ran the pilot cohort up to a hundred and twenty. The pool was wide enough by mid-January. The actual reason we held the gate was that the marketing surface was not ready, which is a worse reason than a calibration shortfall and is the kind of reason we want to write down so we do not give ourselves that excuse a second time.
We shipped Pro 4-month without a step-down contract for candidates whose plans changed mid-cycle.It took until October to write down the proration rule that gives a candidate the per-day rate of the shorter tier when they step down. The four candidates who hit this gap in the intervening months were handled in the co-founders' inbox, case by case, on a policy we made up in the moment. The case-by-case version worked at four candidates. It was wrong policy at the cohort size we were already operating at.
We almost shipped the wrong thumbnail system.Three iterations of the newsroom bento art were so weak a member of the team called them embarrassing. The fourth and current iteration arrived from the design process we should have used from the start. The lesson is small and concrete: the bar is the Apple Newsroom, not the SaaS-blog-of-the-week, and the right people to say so were saying so the entire time.
What we know now
The calibration pipeline is the entire moat. We wrote in December that we were spending more engineering hours on the pipeline than on the loop, and that the loop is seven lines. Twelve months on, the loop is forty-three lines and the pipeline is four hundred. The ratio confirmed itself. Every advantage we have over the category is a calibration-pipeline advantage in disguise.
Pacing is the second moat. The April change that put pacing inside the selection loop was a single sprint of engineering work and produced a measurably better mock score for the same candidate in less of their time. That trade — sharper measurement in less wall-clock — is the trade the candidate is paying us to find, whether they know they are or not. Time is the asset the engine has to spend with respect; we are still the only consumer prep product that treats it that way.
Write the contract before the cohort outgrows the founders' inbox. We learned this the hard way. The November guarantee is the version of that lesson written down. The terms are public, the eligibility is public, the remedy is public. Every future tier ships with its terms written before the surface ships, not after the first candidate calls in.
What 2025 looks like
Five things, planned in the order we expect to ship them, none of them yet on the public surface.
The mastery vector becomes a first-class part of the selection loop. The thirty-element sub- skill structure we stubbed in March now has enough responses against it to drive routing. The post will be called "mastery-vector v2" and will be the next Engineering deep-dive.
Practice mode. A surface separate from the diagnostic and separate from the mock — open-ended adaptive sessions a candidate can run any time, against the engine, without the test-day frame. Some candidates have asked for it since March; we have wanted to ship it since the pacing change worked.
The roadmap. A study plan that the engine writes one day at a time, sequenced against a target date. This was the third bullet in the Pro-Six-Month launch post — what is explicitly not in Pro today. We will ship it when it is ready, not before.
Section Analytics. The retrospective surface — per-section percentile drift, time-per- question shape, the prediction band on the next mock — in alpha through the spring, then beta. The November guarantee post hints at the surface; the spring post will introduce it.
A higher tier with an absolute score floor. Hinted at in the November guarantee, not announced. We will say more in the spring when the work that earns it is closer to shipping.
To the cohort that read every post
We see the read rate on the newsroom. There is a small group of you who have read every post in the order they shipped — the engineering essays the morning they landed, the briefings the day GMAC moved, the tier announcements within an hour. You are not a metric we track; you are the audience the newsroom has been written for. Year One was easier because you were reading. The year-end note exists because you are reading it.
We will keep writing.