A CBSE Class 10 science teacher used to spend two evenings building a single pre-test paper — picking questions, balancing the mark distribution, writing the answer key. On PenLeap, our in-house exam-prep product, she now picks chapters and a difficulty mix and gets a blueprint-conformant pre-test in about 90 seconds, with the marking scheme attached. This is the generation pipeline, the CBSE blueprint rules we encoded, and the quality gate that stops a malformed paper from ever reaching her.
90 sec
Pick to Finished Paper
80
Marks Per Theory Paper
50/30/20
Competency / Constructed / MCQ Split
5
Quality Gates Before Release
## The Answer in 60 Words
The teacher selects chapters, a difficulty mix, and total marks. PenLeap pulls questions from a verified, board-tagged item bank, arranges them to match the CBSE blueprint's mark and question-type distribution, generates fresh variants where the bank is thin, and runs a 5-gate quality check. The output is a print-ready paper plus a marking scheme — produced in about 90 seconds.
## Why This Matters Now
CBSE Class 10 theory papers carry [80 marks, with internal assessment making up the other 20](https://www.businessofapps.com/data/education-app-benchmarks/), and the current pattern is [50% competency-focused questions, 30% short and long answer, and 20% MCQ](https://oswalpublishers.com/blog/cbse-class-10-blueprint/). A pre-test that does not match this distribution does not prepare a student for the real thing — it trains them on the wrong shape of paper. Teachers building these by hand cannot keep the proportions exact across dozens of papers a term. The blueprint is a constraint problem, and constraint problems are exactly what software should own.
## The Generation Pipeline (Pick to Paper)
📋
Teacher picks chapters + difficulty + marks
🗂️
Retrieve from board-tagged item bank
📐
Solve the blueprint constraint
✍️
Generate variants for thin chapters
🖨️
Print-ready paper + marking scheme
The key design choice is the same one behind all our edtech work: questions come from a verified bank first. We only generate fresh items where the bank lacks coverage for a chosen chapter and difficulty, and even then those generated items pass the same gates. This keeps the hallucination surface tiny, the way our
CBSE Class 10 pre-board question generator does for full mock papers.
## How We Encoded the CBSE Blueprint as a Constraint
The blueprint is not a suggestion — it is a set of hard ratios. We turned it into a constraint solver that fills the paper to spec.
🎯
Question-type ratios
50% competency-based (case/source/MCQ), 30% short and long constructed-response, 20% select-response MCQ. The solver will not return a paper that misses these bands.
⚖️
Mark distribution
Total marks hit the teacher's target (default 80). 1, 2, 3 and 5-mark questions are balanced so the paper sums exactly and matches typical section weights.
📚
Chapter coverage
Marks are spread across the selected chapters in proportion to their syllabus weight, so no single unit dominates a "balanced" pre-test.
🔀
Difficulty mix
Easy / medium / hard ratios follow the teacher's slider. The solver treats difficulty as another dimension it must satisfy alongside type and marks.
Why a solver, not a prompt: we tried asking a model to "make a balanced 80-mark paper." It produced papers that summed to 78 or 83 and quietly skewed toward MCQs because they are easier to generate. A constraint solver guarantees the arithmetic and the ratios. The model only writes the prose of individual questions, never the paper structure.
## The 5 Quality Gates
Every paper passes five checks before a teacher sees it. Any failure sends the item back for regeneration or human review.
1
Marks sum exactly
The paper must total the requested marks, with section sub-totals correct. An off-by-one here is the most common and most embarrassing failure, so it is gate one.
2
Blueprint conformance
The 50/30/20 type split and difficulty mix are re-verified on the assembled paper, not just assumed from the solver.
3
Answer-key integrity
Every question has a marking scheme. For MCQs, exactly one option is correct. For constructed-response, the rubric and mark allocation are present and sum to the question's marks.
4
Duplicate suppression
No two questions on the paper test the identical concept with near-identical wording, and the paper is checked against recent papers the same teacher generated.
5
Curriculum-scope check
Every question maps to a selected chapter and the Class 10 syllabus. Out-of-scope or out-of-class content is rejected — a real, correct physics question from Class 11 is still wrong here.
## How We Hit 90 Seconds
Speed is mostly about not generating what you can retrieve.
- Retrieval from the item bank is near-instant; generation is the slow part, so we generate as little as possible.
- The constraint solver runs in milliseconds — it is arranging existing items, not writing them.
- Variant generation for thin chapters runs in parallel, not sequentially, across the questions that need it.
- The marking scheme for retrieved questions is pre-stored, so only generated items need a fresh key.
- Quality gates run as fast deterministic checks first, escalating to model-based checks only for generated items.
A paper that is 80% retrieved bank questions and 20% freshly generated variants lands in about 90 seconds. An all-generated paper for an obscure chapter combination takes longer — and we show the teacher a progress state rather than pretend it is instant.
## When Not to Auto-Generate
Auto-generation is right for routine pre-tests and practice papers, where the value is speed and blueprint conformance. It is the wrong tool for a high-stakes school exam that will appear on a report card, or for a brand-new chapter where the item bank has no verified coverage and every question would be model-generated. In those cases we route to a human-authored or human-reviewed flow. Speed is only a feature when the floor of correctness is already guaranteed; for the questions that decide a grade, a teacher should still hold the pen.
The gate that earns its keep: gate 3, answer-key integrity, catches the worst real-world failure — a paper that looks perfect but has a wrong or missing answer in the key. A student loses marks for a "wrong" answer that was actually right. We treat a key error as a release-blocking bug, not a warning.
## Real Example: The Two-Evening Paper, Now 90 Seconds
A Class 10 science teacher in a CBSE school told us she budgeted two evenings per pre-test paper — selecting questions, balancing marks, and hand-writing the key. With PenLeap she now generates a blueprint-conformant pre-test during a free period and spends her evenings reviewing student answers instead of authoring papers. The same engine scales up to our full mock-paper pipeline described in
generating 200 custom practice papers in the pre-board pipeline, and the rubric side is detailed in our
Hindi essay-grading engine post. See the product at
PenLeap. This is the kind of constraint-driven generation our
AI and automation team ships, proven in production on PenLeap.
## Frequently Asked Questions
### Does PenLeap make up exam questions with AI?
Mostly no. It retrieves questions from a verified, board-tagged item bank first, and only generates fresh variants where a chosen chapter lacks coverage. Generated items pass the same five quality gates as bank items, including a curriculum-scope and answer-key check.
### How does it guarantee the paper matches the CBSE blueprint?
A constraint solver assembles the paper to hit the exact mark total and the 50% competency / 30% constructed-response / 20% MCQ type split, then a separate gate re-verifies conformance on the finished paper. The model writes individual questions; it never decides paper structure.
### What does the 90 seconds actually include?
From the teacher pressing generate to a print-ready paper plus marking scheme. It is fast because most questions are retrieved, not generated, and only thin-coverage chapters trigger parallel variant generation.
### Can teachers edit the generated paper?
Yes. The output is a starting point a teacher can adjust — swap a question, change marks, edit wording. The blueprint solver and gates ensure the auto-generated version is already conformant, so edits are refinements rather than repairs.
### Does it work for subjects other than science?
The engine is subject-agnostic as long as a board-tagged item bank exists for that subject. We built coverage subject by subject, and the constraint and gate logic is shared. The bottleneck is always the verified item bank, not the generator.
### What happens if a chapter has no questions in the bank?
The system generates fresh variants for that chapter, but flags the paper as containing generated content and applies the strictest gates. For high-stakes use we recommend a human review pass on any paper with significant generated content.
Want a constraint-driven document generator for your domain?
We build blueprint-conformant generation engines — papers, reports, contracts — for Indian teams in 4-7 weeks. Typical cost: ₹7L-₹16L for a production v1 with a verified content bank and quality gates. Suitable if your output must hit strict structural rules every time. No slides — just your spec and our honest take. Reach us at contact@softechinfra.com.
Book a 20-min Call