How our questions are made, checked, and corrected
Nursing-exam prep has a trust problem: invented faculty, invented ranks, invented authority. The only useful answer is to show the process. This is ours, written plainly, including the parts that are limitations.
NurseDrill grew out of the ideas and everyday struggles of a group of nursing professionals who wanted to keep learning while working, without the hassle. That is what it is built for. What you can check for yourself: a written process, source citations counted, not promised (3,245 of 3,353 questions name their source. How we count), and a public record of every error we have fixed.
Questions, corrections, and complaints: contact page.
The pipeline, stage by stage
New questions are commissioned against a blueprint derived from what the exam has actually asked — subject weights and topic frequencies taken from past NORCET Mains papers, not from a wish list. A topic that has appeared in eight of nine papers gets more questions than one that appeared once. We would rather ship 40 questions on the right topics than 400 on the wrong ones.
Every question passes machine validators before anything else happens. They reject: a missing or out-of-range answer key, options that aren't exactly four, a missing explanation, a missing source citation, two options that say the same thing, a catch-all "all of the above" that makes the key ambiguous, and — the useful one — an explanation that contradicts its own answer key. These run on every content change, and a failure blocks the change.
The validators are part of the codebase (scripts/validate-questions.mjs) and run in CI on every content change — a failure blocks the change from shipping.
An item is drafted in one model context. It is then handed to a second, separate context — a fresh Claude conversation that sees only what a candidate sees (the stem, the four options, the keyed answer and the explanations shown on review) plus the citation and the written rubric below, and never the drafting conversation, its source notes or its earlier drafts. That second context is instructed to refute the item, not to agree with it, and returns one of three verdicts — pass, refute or unsure — with its reasons. The item passes only if every point of the rubric holds:
- The keyed answer is correct against the cited source and current national guidance
- Every distractor is a defensible alternative, none is trivially wrong or a second correct answer
- An image item cannot be answered without its image; no item is answerable from option shape alone
- The explanation and per-option explanations agree with the key
- The citation names the chapter or guideline that actually contains the fact
- The fact is current national guidance, not a superseded value
- No option-shape cue: the key does not stand out by length, detail or hedging, and no distractor is a throwaway
- Every distractor is a belief a nurse could actually hold
- The explanation teaches this item: the fact that decides it and why, not a line that would fit any question
- Where reasons are given for each option, each is specific to that option, not a template repeated across them
The verdict is recorded against the item with the method and the date. Only an item with a recorded pass carries the "verified" label and its date on the source line; an item without a record shows no label. Any model family may run either pass — what makes the check independent is the separation of contexts and the written rubric, not the vendor. The verifier tool is itself tested against a fixture of deliberately broken items it has to catch.
The rubric above is read from the same file the verifier embeds in its prompt (data/mains/verification-protocol.json), so this page cannot describe a rubric the tool does not use.
A refute or an unsure verdict removes the item's verification record, so it shows no label. The item is then either corrected and put through the adversarial pass again in a new context, or dropped. Nothing flagged ships as verified, and no human clinical review is claimed anywhere in this process.
Every correction to content that has already shipped gets an entry in the public fix-log with the date, what was wrong and what changed. Where a fact is genuinely contested between guidelines (they exist — resuscitation drug timing is the classic), the explanation says so instead of pretending there is one answer.
2,813 of 3,122 servable items carry a verification record under this protocol (90%), counted from the question files when this site was built. That count is the Mains bank and the free topic banks; every one of the 607 previous-year items carries one; every one of the 210 drill items carries one. Every new item is verified before it ships. Items without a record show no label until they pass.
Every question names the reference that governs the fact it tests — a standard textbook chapter or a national guideline: Park for community health, Brunner & Suddarth for medical-surgical, KDT for pharmacology, Dutta for obstetrics, Ghai for paediatrics, the MoHFW rules and programme documents where the fact is Indian policy. This is enforced by a validator, so it cannot quietly lapse.
The limitation, stated plainly: citations are at chapter or guideline level, not page level. We would rather give you a reference you can verify than a page number we invented.
Every question in the product has a report control. Reports are triaged weekly; fixes ship with the weekly publish, and each one is written into the public fix-log with a date, what was wrong, and what changed. Nothing is quietly edited. If we got something wrong and you told us, the record of that is public.
The newest corrections in the fix-log:
- 31 October 2026 · , 100137, Mission Indradhanush catch-up scenario
- 31 October 2026 · Question-bank, item 2367, oral iron — how to take it (pregnancy)
- 31 October 2026 · Chn 100157, India HDI rank
The works those citations name, with every question that cites each chapter: the sources library. Its “How we cite” block is the citation stage above, word for word.
Previous-year papers
Every previous-year set on NurseDrill is memory-based recall: questions reconstructed from what candidates remember, reworded by us, with explanations we wrote. We do not reproduce official question papers — they are copyrighted, and we won't build a product on someone else's paper. Where a recall set is incomplete or a remembered option looks wrong, the explanation says so.
Practical consequence: a recall set is excellent for pattern and pressure, and imperfect as a source of truth. Treat the explanation and citation as the authority, not the recalled option.
One more provenance note: the topic-frequency tiers on the syllabus tracker are transcribed from images circulated in a NORCET toppers' community (WhatsApp, April 2026). The source images aren't independently verifiable, so we present those tiers as community signal, not systematic analysis.
The weekly rhythm
During exam season a new full mock ships each week. It goes through every stage above. If a mock fails its gates, it ships late rather than unreviewed. Off-season the cadence is slower, and we'd rather print that than pretend otherwise.
Support runs on printed limits too: replies within 48 hours, payment and access problems handled the same day, fixes batched weekly. Details on the contact page.