Year One ended with a list and a forecast. The list was what had shipped in 2024; the forecast was a five-point sketch of 2025 — mastery vector v2, Practice, the Roadmap, Section Analytics, a higher tier with a longer commitment. All five shipped. Two of them shipped roughly when the forecast said they would. The other three slipped by a quarter and arrived in a different order than the letter named. This is the year-end note on what that produced.
We are writing this on a quieter day than the one we wrote Year One on. There were roughly twenty-one hundred candidates in the engine when last year's letter shipped. Today there are roughly seven thousand four hundred. The product has more surfaces, the team has more people, and the cohort has more shape — but the work in front of us still reads the same: build the version of GMAT preparation a serious candidate would actually want to use.
What shipped, in order
Each of these was its own note in the newsroom. The list, again for the record:
January. Mastery vector v2 — the thirty-element sub-skill structure becomes a first-class selection input. Data Insights gets the three-axis decomposition (interpretation, inference, integration) the section needs. February. Practice — open-ended adaptive sessions, the third engine surface after the diagnostic and the Mock. March. The Roadmap — a per-day study plan the engine writes, rewritten every night against the candidate's vector and target date. April. Section Analytics in private alpha. May. Ultra at $1,599 / six months with the absolute 715+ floor — the same six-month-renewal remedy as Pro, at a higher bar. June. The scratchpad inside the Mock runner. July. Library v0 — the first five interactive lessons, Number Properties. August. The annual GMAC percentile refresh — half the size of the 2024 shift, as forecast. September. Library v0.5 — five Critical Reasoning Assumption lessons, argument-first. October. Section Mode promoted from a thin Mock wrapper to its own first-class surface. November. Section-aware calculator inside the runner.
What we got wrong
Three things, in keeping with last year's letter. The absence of a "what we got wrong" section in a year-end note is the surest sign the writer is avoiding the year.
Section Analytics took longer to go to beta than we said it would. The forecast in Year One was "in alpha through the spring, then beta." Section Analytics is still in a fifty-candidate alpha as of October. The reason is the one we wrote into the April launch post — the Monte Carlo run behind the prediction band was fast at fifty candidates and not fast at four thousand, and the engineering work to make it cheap at scale ran behind the rest of the year's roadmap. The beta will open in 2026. We did not write a beta date into this letter.
The Library Quant-foundation timeline slipped a quarter. We had said in July that two further Quant batches would land in the autumn — fractions/percents and ratios/proportions. The first landed; the second slipped into the new year. The cause was authoring throughput rather than pedagogical disagreement; once the format was right on NP and CR, the bottleneck moved to writing speed, not slide-design judgment. The fix is the v1 authoring toolchain, in build now.
The formal guarantee policy did not get written until the spring. We have been operating against the November 2024 contract since last year and the May 2025 Ultra addendum since Ultra launched. Both work. Neither is the single consolidated document the candidate ought to be able to read, at one URL, before they pay. That document is in draft now and ships in April under the Policy badge.
What we know now
Three readings on the year, less abstract than that opener.
The contract held. The November 2024 guarantee got its first real test in May 2025, when the earliest Pro 6-month cohort exited the window. Of the candidates who completed eligibility and missed their banded lift, the renewal remedy triggered exactly the way the contract said it would, for exactly the candidates the contract said it would. Ultra had its first cohort close in November; two candidates who cleared eligibility and missed the 715+ floor had their Ultra windows renewed at no charge inside the week the live score landed. We will publish a longer note on the renewal-versus-churn pattern of the first Ultra cohort in January, when the data is a quarter old and worth a separate post.
Faithfulness is not optional. The November calculator-rules change is the year's clearest case. The candidate, given the choice between an easier rehearsal and a more faithful one, will almost always pick the easier one in the moment and regret the choice in the room. The right answer is not to ask. The Mock surface does not let the candidate keep a calculator open on Quant the way the scratchpad's first six months did, and the surface is less popular for it on the day of release and more useful for it on test day. We will keep making that trade.
The slide grammar generalises slowly. The Library's argument-first frame works on Critical Reasoning the way the rule-derivation frame worked on Number Properties. Both lessons taught their topic more efficiently than the textbook alternatives. What the data is also telling us is that Quant and Verbal do not authorise at the same speed; the Verbal batch in September took less calendar time than its per-slide complexity should have suggested, while the planned ratios-and-proportions batch took more. We are learning slower than the optimistic version of the roadmap assumed. The Library will be very good. It will not be very fast.
What 2026 looks like
Five forecasts, again. The track record on last year's five is documented above; treat this year's accordingly.
The Library completes its Quant foundation and opens the Verbal section properly. Fractions/percents and ratios/proportions ship in the late winter; further CR families queue behind the Assumption batch through the spring. Data Insights and Reading Comprehension are the longer-horizon sections; each is in pedagogical research, neither has a launch date yet. The Library Vision essay in May will write down the long-range target — roughly two hundred lessons covering the whole test — alongside what is actually live today.
Section Analytics goes to broader beta. The Monte Carlo engineering work is on the January sprint. The private-beta cohort widens from fifty to the full Pro and Ultra opt-in surface; the surface formally goes "beta" rather than alpha. The four views the original spec named — predicted band, per-section percentile, time-per-question drift, and topic-mastery deltas — get whichever subset the alpha cohort tells us is worth keeping. We expect to cut one.
The guarantee, written down in one place. The November 2024 contract and the May 2025 Ultra addendum get consolidated into one canonical document candidates can read at /guarantee before they pay. Same terms, one URL. Spring ship.
A second Letter. The first numbered Letter from this account was the founding brief in September 2023; the year-end letters have run since. The next standalone Letter — not a year-end note — is in draft now. It is about what serious GMAT preparation actually requires and why most of the industry cannot honestly offer it. It will run in the spring under Nicola's signature alone.
The longer engine essay. We have wanted to write a single deep technical post on how the adaptive loop actually decides what to ask next — not a launch post, not a release note, the canonical piece — since the v1 launch in March 2024. The piece was always one quarter from being right. It is on the spring's editorial list.
To the cohort that read every post, again
The same group of you who finished last year's letter finished every post that has shipped since. We see the read rate; the rest of the team sees it too. The newsroom is built for you. The product is being built next to it.
Two years in. We will keep writing.