About tensor.news
tensor.news is the evidence layer for AI model claims and benchmarks. We answer one question the leaderboards skip: which scores are trustworthy, comparable, and worth acting on.
Why we exist
A benchmark score is a function of the model, the harness, and the protocol, not of the model alone. Run the same model through a different scaffold and the number moves. Most coverage reports a single figure without saying whether the benchmark still separates models, whether the score was independently reproduced, or whether the test set could have leaked into training. We make those distinctions the product.
What we do
- Benchmark Integrity Index. We grade every benchmark from 0 to 100 on discrimination, saturation, contamination risk, harness comparability, and age, so you know which evaluations still mean something.
- Model Claim Ledger. Every model-by-benchmark score is tagged by evidence status: independently reproduced, self-reported, contradicted, or unverified. We keep independent evaluations and vendor self-reports separate rather than averaging them.
Who runs it
tensor.news is an independent research desk in Istanbul. We are not affiliated with any lab, vendor, or benchmark maintainer, and we do not run sponsored placements. The record is kept by an automated pipeline that compiles public evaluation data and narrates only from cited facts; every page carries its sources in the provenance panel.
Standards
- Provenance. Every fact traces to a source record: which evaluator ran the score, when it was measured, and where it was published.
- Honesty about what a number is. We frame every score as task performance under a disclosed harness, never as deployed, real-world, or agentic capability.
- Independence. Where a benchmark's integrity is low or its harness is mixed, we say so next to the score rather than leaving it to be read as a verdict.
Sources
Frontier benchmark scores are drawn from independent evaluations, currently anchored on Epoch AI's capabilities data, alongside vendor technical reports (labeled as self-reported). Model metadata and open-weight releases come from Hugging Face; paper records from OpenAlex; pricing from LiteLLM and models.dev. The data refreshes daily. Read the full grading method on the methodology page.
Corrections
The record is not infallible, but it is checkable. If a fact looks wrong, tell us — we verify against the cited source and fix or annotate the entry. See the contact page for how corrections work and the response time you should expect.
Contact
Have feedback, a correction, or a tip? Reach us at hello@tensor.news or use the contact page.