How we evaluate your UPSC Mains answers
The per-sub-part rubric, the question-specific scoring pipeline, the OCR step for handwritten answer sheets, what an actual scored answer looks like — and the things we deliberately don't claim.
Last updated 2026-08-25 · Authored by Rahul Choudhary, Founder
In one paragraph
Every question on this site has a rubric attached to it before any answer is graded, and it is built from the question as printed. Each sub-part gets its own block carrying the marks and word limit read off the scanned paper, the command word that fixes the shape of a correct answer, the three or four things that must be covered, what earns more, and the traps that lose marks. A five-part question is marked five times rather than judged once as a whole. You receive a score, dimension-level rationales, missing keywords/examples, and a concrete next-step suggestion. Handwritten answers pass through OCR first; mixed-script (English + Devanagari) is supported.
Built from the question as printed
A rubric is written for one question, never stamped from a template. Marks and word limits come from the paper itself; the rest is what that question, in that subject, actually rewards.
1. One block per sub-part
A question with parts (a) to (e) gets five blocks, not one impression of the whole. Each carries the marks and the word limit printed on that sub-part, read off the scanned paper rather than assumed — so five sub-parts worth 10 marks each are marked out of 10 five times, not out of 50 once.
2. The command word, per sub-part
Most marks lost on a question the candidate knew are lost by writing the wrong shape of answer. Each sub-part names its own directive — a five-part question usually has five — with the structure that directive requires and the mistake it invites. "Comment" wants a position; "examine" wants reasoning carried by evidence; narrating facts under either loses marks.
3. Must cover
Three or four non-negotiables for that sub-part, each a phrase rather than a sentence. Missing one caps the sub-part below half marks. They are concrete — the scheme, the species, the statute, the theorem, the thinker — because "shows good understanding" tells an examiner nothing.
4. Earns more, and the extra mark
What lifts a correct answer: a named committee or judgment, a precise figure, a diagram worth thirty seconds, a current-affairs anchor. Held apart from the non-negotiables on purpose, so an answer is never marked down for omitting something the paper did not ask for.
5. Loses marks
The traps for this sub-part, including the directive's own characteristic failure — MSP-only solutions on an agriculture question, dramatic writing on internal security, emotive language on a society question, a final value with no working on a derivation.
What changes by paper
Each paper is marked the way that paper is actually written. These are the same technique cards the answering guide shows you.
- →Mathematics, Statistics, Physics, Engineering: Marked on method — assumptions stated, each step justified, units carried. A correct final value with no working is not a complete answer.
- →Chemistry, Geology: Mechanisms and structures drawn, correct IUPAC naming and stereochemistry, the named Indian formation where the question is regional
- →Botany, Zoology, Agriculture, Animal Husbandry: Exact binomials, labelled diagrams, the pathway or clinical sequence in order
- →Sociology, Anthropology, Psychology: The thinker named before the idea, one ethnographic or empirical case, Indian context in Paper II
- →History: A thesis rather than a narrative, sources cited, the historiographical school and its rival named
- →Law: The section or Article precisely, the case and its ratio rather than its outcome, law applied to the given facts
- →GS-4 Ethics: The crux, stakeholders including the silent ones, the dilemma named as competing duties, options weighed, action in three horizons
- →Essay: Decode before writing, a thesis held throughout, the quote returned to in every section
From your answer to your score
Five stages. Average end-to-end latency: 25–30 seconds for typed answers, 45–60 seconds for handwritten.
- 1
Question selection
You either paste a question from anywhere or pick one of our 2,400 PYQs from UPSC Mains 2021–2025. Each PYQ row in our database carries the original UPSC text (English + Hindi), marks, word limit, paper number, and topic tag.
- 2
Rubric pre-generation
Before scoring your answer, the system generates a question-specific rubric — 5 weighted dimensions, expected key-points list, criteria for poor / average / excellent answers on each dimension. The rubric was already generated and stored for every PYQ in our bank using a separate prompt run (Kimi K2.5 on AWS Bedrock, mechanical-validator gated). For pasted custom questions, the rubric is generated on the fly.
- 3
OCR (handwritten only)
Photographed answer sheets pass through a layout-aware OCR pipeline that handles Devanagari and English mixed scripts and preserves diagram positions. Confidence scores are surfaced — if the OCR drops below 70% confidence on any sentence, we flag it for you to review before scoring.
- 4
Rubric-guided scoring
The grader uses the rubric as its primary scoring guide — checking your answer against each dimension's key points and excellent/average/poor criteria — and reports four parameter scores on 0–5 you can track across every evaluation: Content, Structure, Analysis, Language. Scores are calibrated: a 4/5 means "consistently above-average"; 5/5 is reserved for answers that show analytical lift beyond the model.
- 5
Aggregate score + feedback
Marks are awarded against the question's real UPSC maximum, informed by the rubric's weighted dimensions. You get the score, the four parameter scores, what was strong, what was missing (specific keywords / examples / cases), and a forward-looking suggestion — "to move from 6.5/10 to 8/10, address X, Y, Z."
What a scored answer actually looks like
"The growth of cabinet system has practically resulted in the marginalisation of the parliamentary supremacy." Elucidate.
Treated "elucidate" correctly as explanatory-demonstrative, not evaluative. Cited Article 74 + 75 + 123. Mentioned Kihoto Hollohan (1992) on Speaker's role.
Add: 91st Amendment (2003) anti-defection example; 2020 farm-laws ordinance precedent; Ivor Jennings' framing in conclusion. The omission of the Jennings reference is the single biggest reason this didn't score 8+.
Anti-defection · 91st Amendment · NITI Aayog · Ivor Jennings · DRSCs (Departmentally Related Standing Committees)
Things we deliberately don't claim
Honesty here matters more than marketing copy.
We do not predict your actual UPSC marks. UPSC is a comparative exam graded by humans against unpublished bell-curves; no AI tool can predict that. Our scores correlate with examiner feedback patterns but should be read as "improvement signal," not "expected mark."
For deeply technical Optional sub-parts that hinge on a single derivation step or a specific equation, our grader can miss subtle errors a domain expert would catch. Mathematics and Engineering answers should be cross-checked against textbooks for the worked-step accuracy.
For Hindi-medium answers, our rubric and feedback are generated in Hindi when language=hindi is selected, but the model's feedback specificity in Hindi is currently ~80% of its English specificity. We are working on closing this gap.
Current-affairs cutoff: examples cited in our model answers go up to early 2026. Events after that may not be reflected until the next content update.
Who reviews the rubrics
Rahul Choudhary
Founder & editorial lead
IIT Delhi alum; built and maintains the rubric pipeline and PYQ corpus.
IndiaSchool.ai editorial board
Subject reviewers
Content reviewed periodically against published UPSC topper answer copies and coaching-institute model answers.
Operated by IndiaSchool.ai Pvt Ltd. Read more about the team →
Try it on a real PYQ
2,480+ questions, GS + Essay + 25 Optionals. Five free evaluations every month.