About

Model Report Card

AI benchmarks taken apart: who ran it, how it was scored, and what the number means.

Benchmarks Curated publishes technical teardowns of AI benchmarks and evaluations across LLMs, agents, voice, video and physical AI. For every benchmark we cover who runs it, how it's scored, sample size, version and contamination risk, and whether a score is lab-reported or independently measured. Disclosure: affiliated with Humanbit, a data labelling and evaluation company.

How every story is checked

1. Research starts from primary sources: filings, company releases, official documentation, papers and government data.
2. Every person, company, product, place and number needs a primary source or two independent ones before it can appear.
3. A separate fact-check pass, which did not write the draft, re-checks each name and number against its source. A story that fails is held, not published.
4. Prices, salaries, scores and rankings carry an "as of" date. Company and lab claims are labelled as theirs.
5. We never claim hands-on testing we did not do.

Research, drafting and fact-checking are assisted by AI tools working under these rules. Sources are listed at the end of every story.

Corrections

When we get something wrong we fix it and add a dated correction note to the top of the story, saying what changed.

Disclosure

Disclosure: this site is affiliated with Humanbit, a data labelling and AI evaluation company.

Who writes here

• Shivam Singh · LinkedIn