We tested our AI against real UPSC examiners. Here are the results.

Each dot = one real answer · dashed line = perfect agreement

What this shows — honestly

• Our AI never inflates. Across the test set the AI averaged 38.3% vs the human examiners’ 46.3% — when it errs, it errs strict. The score you see is a floor to build from, not flattery.
• The median difference is zero. For the typical answer, the AI lands on the same mark the human examiner gave.
• Human examiners disagree too. Studies of UPSC-style essay marking show human-to-human variance of several marks on the same copy. Our published spread (50.8% within ±1 mark, 54.1% within 10 percentage points) is measured against a single examiner’s judgment — the hardest possible test.
• This page will keep updating. As our evaluated-copy corpus grows, we re-run this benchmark and publish the new numbers — good or bad.

Methodology

1. We collected 61 handwritten mains answers that were already evaluated and scored by experienced human examiners (mock-test copies of real aspirants, including future toppers).
2. Our AI evaluator scored the same handwritten pages with its standard 9-parameter UPSC rubric, instructed to ignore any visible examiner markings.
3. We compared the two scores per answer. No answers were excluded for scoring badly; the full set is plotted above.
4. Benchmark last run: 2026-07-13. Model and rubric identical to what evaluates your answers.