At one point, ScholarBank could advertise a catalogue of 15,204 multiple-choice questions. Fourteen thousand four hundred of them came from a deterministic expansion system. The number looked substantial, the catalogue could be searched, and the software could verify that every record had the fields it needed. I removed that expanded core anyway.
This was not a technical failure. It was a product decision about what a question count is allowed to imply. A large bank suggests variety, judgement, and a depth of practice that the number alone cannot prove. If I could not defend those implications, keeping the number would make the interface more confident than the evidence behind it.
A validator is not a classroom
Structural validation is useful. ScholarBank's checks can catch duplicate identifiers, malformed options, missing explanations, impossible answer indexes, broken classifications, and catalogue totals that drift away from the source. They can confirm that a route returns the intended question and that protected content is not exposed through a public catalogue response. Those are real safeguards.
They are not calibration. A passing schema cannot tell me whether a question is genuinely easy for one school-year cohort and difficult for another. It cannot show that a distractor is plausible, that a passage demands the intended skill, or that a timing estimate survives contact with learners. Those claims need independent review, pilot item-analysis, accessibility and rights checks, versioned evidence, and a correction process. Deterministic expansion made consistent records; it did not create that evidence.
A smaller number with a clearer boundary
The current product defines 804 released multiple-choice questions, alongside 12 separate, timed writing tasks. I keep the writing tasks outside the multiple-choice count because writing is not another answer bubble. It needs a prompt, independent composition, and a different review ritual. The remaining catalogue is still not an official provider paper, a scholarship-outcome guarantee, or an empirically calibrated simulation.
I also separated two pieces of information that preparation products often blur: the student's current school year and the year they hope to enter. The target entry year belongs to planning—school dates, provider context, and how far backwards to work. Practice difficulty should begin with the learner's current year. A student planning for a later entry does not suddenly acquire the knowledge of that later cohort. ScholarBank can remember both facts without pretending they mean the same thing.
Make the learning loop trustworthy
Removing the expanded catalogue shifted my attention from inventory to delivery. Protected questions are delivered through a server-side session rather than poured into the browser. The server checks the requested scope and access, and correctness is recalculated from the canonical bank instead of trusting a score submitted by the client.
Feedback is deliberately deferred until the whole set is complete. In-progress session reads omit unseen content and answer keys; only a completed review can include the worked explanation and reusable strategy. That choice protects the bank from being harvested one answer at a time, but it also preserves the continuity of an attempt. The learner finishes the set before the interface changes from asking to teaching.
Progress without a prediction costume
I simplified the default progress model as well. One account can begin practising and build one history of answers, pace, sessions, and streaks without first constructing a family hierarchy or selecting a learner profile. The dashboard can use that history to suggest a weaker subject or a timed set. That is a practical next step, not a claim to know a student's future.
Even the headline practice indicator is rule-based. After enough activity for the interface to show it, a fixed formula weights recorded accuracy and adds a bounded amount for the size of the evidence base. It is not an AI model, percentile estimate, provider score, or scholarship prediction. Before 20 answered questions, ScholarBank withholds the number and asks for a broader baseline instead. Naming the indicator modestly is part of making it honest.
What I still do not know
A narrower release does not resolve the hard questions. I do not yet have independent evidence that every difficulty label is right, that the collection is balanced for real cohorts, or that the recommendations improve outcomes. The separate editorial draft pack remains outside the learner-facing release until its review gates are satisfied. Billing foundations exist in the code, but the current beta should not be described as an operating paid service.
Removing 14,400 questions made ScholarBank numerically smaller and conceptually stronger. The useful question is no longer “How large can the bank look?” It is “What can each part of the product truthfully claim?” For now, 804 questions, 12 writing tasks, a guarded feedback loop, and visible uncertainty are a better answer than five digits of unearned confidence.