What is Bharat Score?
// definition
Bharat Score is the macro-mean accuracy a language model achieves across the seven categories of the BKP-500 benchmark. Each category's accuracy is computed first, then the seven category scores are averaged with equal weight.
It is reported as a percentage, with a 95% confidence interval.
// how it is computed
1. Score every India-specific item in the corpus as correct or incorrect, using the deterministic grader declared for that item's answer type.
2. Compute accuracy separately within each of the seven categories.
3. Take the unweighted mean of those seven category accuracies.
In formula terms, with ai as the accuracy in category i:
// why macro-averaged rather than a raw item mean
The seven categories do not contain equal numbers of items. A raw mean across all items would let the largest category dominate the headline figure, so a model that happened to be strong in the biggest category could post a high score while being blind in several others.
Macro-averaging removes that. Under this definition a model that scores 78 on Indian numerals and near zero on land units cannot hide the blind spot behind item volume. Every category counts for one seventh of the score regardless of how many items it holds.
The practical consequence: Bharat Score is harder to achieve than a raw accuracy figure would be, and the two are not interchangeable. A model at 45% Bharat Score is not answering 45% of items correctly.
// the seven categories
Indian numeral system · Weights, volumes and informal measures · Land units by state · Agricultural seasons and crop calendars · Fiscal-year conventions · Government schemes · Structural identifiers.
Each is defined in full on the BKP-500 benchmark page.
// how to read the confidence interval
Every Bharat Score is published with a 95% interval, bootstrapped over one value per item rather than per raw response row. Repeated samples of the same item are never treated as independent observations.
Overlapping intervals are not a difference. In the 2026 run the intervals overlap across much of the roster, so a model ranked 6th and a model ranked 8th may not be distinguishable. Adjacent-rank comparisons are tested with Holm-Bonferroni correction at a 0.05 family-wise alpha, and only the corrected result supports a claim that two models differ.
// related
Locale Gap Delta is the companion metric, measuring the difference between a model's performance on India-framed items and their matched international controls.
// where the term comes from
Bharat Score was defined by Sthānika AI for the BKP-500 benchmark. It is not a pre-existing measure adopted from other work.
// citation
Sthānika AI (2026). BKP-500: the Bharat Knowledge Probe. https://sthanika.ai/benchmarks/bkp-500