The problem with building a bias analysis tool is that you're asking people to trust your judgement about trust. That felt like a problem worth solving properly before Lenstrum went public. So we built a corpus.

A corpus — in this context — is a curated dataset of content items where you already know the right answer. You feed the tool your hundred items, compare its outputs against the known results, and see how often it gets it right. It's the most honest way to answer the question that matters: is this thing actually working? Most bias analysis tools don't publish accuracy figures. I understand why — it's a hard problem, and the numbers are uncomfortable to show publicly when they're imperfect. But if Lenstrum is asking users to act on its outputs, hiding the accuracy data seems like exactly the wrong move. So we're publishing ours.

We built a 100-item corpus spanning five categories: 40 political articles, 25 political YouTube videos, 15 non-political YouTube videos, 10 non-political articles, and 10 centre or balanced sources. For validation, we used AllSides and Ad Fontes Media ratings wherever they covered the content. For content types with no published third-party benchmark — which is a lot of YouTube, and most non-political content — we used internal reference ratings. That's a real limitation and we're honest about it. Internal ratings introduce the same subjectivity we're trying to guard against. But the alternative is to only test on easy cases, and that seemed worse.

The headline numbers: 90% overall accuracy. 100% on political YouTube. 95% on political articles. Those are the categories where Lenstrum performs strongest, and they're also the ones with the most reliable third-party benchmarks to validate against. When AllSides says a source leans right and Lenstrum says the same piece leans right, that's the kind of convergence we were looking for.

Non-political content tells a different story. For theology, economics, philosophy, and science, Lenstrum is scoring in the 73–80% range. That's not surprising — these are genuinely harder categories, the framing signals are more subtle, and there's no published third-party benchmark to validate against. We think directional accuracy is still valuable here, but we're not going to overclaim. The confidence score is Lenstrum's built-in way of flagging uncertainty: when it returns Low confidence, that's the tool telling you to apply more scepticism.

The finding that gave us real confidence was this: high-confidence outputs — the ones where Lenstrum says it's sure — match at 92.9%. When Lenstrum is confident, it's right the vast majority of the time. The confidence level isn't decoration. It's a meaningful signal about output quality.

There's a structural limitation in how we validated that's worth being explicit about. AllSides and Ad Fontes rate outlets and channels — not individual articles or videos. Lenstrum rates individual pieces of content. A single outlet can produce content rated differently on the same day, depending on the specific piece. So when we say 90% accuracy, we don't mean that 10% of analyses are simply wrong. We mean the methodology has edges — places where individual content deviates from outlet-level benchmarks — and the confidence score is designed to signal where those edges are.

What comes next: we're expanding the corpus to 200 items post-launch, using real URLs submitted by users. The thumbs up/down feedback on every result will feed into a separate validation signal. After any significant prompt change, we'll re-run the corpus and publish updated figures. The corpus is a starting point, not a finish line.

We're not claiming perfection. We're claiming we checked, we published what we found, and we'll keep checking. In a space where most tools don't show their work, that feels like the right place to start.

The full methodology is published at lenstrum.com/methodology. If you want to try Lenstrum for yourself, start here.