Since its inception, every DNA45 answer has carried a Confidence Score. This score is the system’s honest read on how strongly an answer is backed by credible external sources at the time of publishing. Sources from established institutions tend to contribute to higher scores due to their consistent citation and verification across independent references.

In a time of widespread automation and content generation, it has become increasingly difficult to rely on unverified sources — such as auto-generated blogs, transient social media posts, or newly created websites that may lack credibility or longevity.

The Confidence Score is how DNA45 answers that problem in public — by showing, on the face of every card, how much of the answer rests on sources that have been independently corroborated.

To illustrate how this evaluation system works:

Every answer DNA45 produces is shipped with a single number, written like this:

It is not a grade. It is not a self-assessment. It is not self-conviction. It is a public-facing measure of how strongly the claim in the card is supported by credible, external evidence at the moment it is rendered.

Lower scores do not imply the claim is incorrect. They indicate that, at the time of evaluation, there is limited corroborating evidence from independent, credible sources. This preserves evidential neutrality by separating what the model can produce from what can be externally verified. It allows users to interpret each claim in proportion to its available evidence and remain aware of where verification is still evolving.

What the number actually means

The score is the answer to one question:

If you set out to verify this claim right now, how much external, trustworthy material would you find pointing in the same direction?

A high score means a lot. A low score means very little. That is the whole framing.

How to read the bands

• 80–100% — Strongly backed. Multiple top-tier sources independently support the claim. Treat as a settled position you can act on.
• 60–79% — Well backed. A handful of credible sources line up. Treat as a strong working assumption; one extra check before high-stakes decisions is wise.
• 40–59% — Partially backed. Some external evidence exists, but the claim leans on the model’s training as much as on outside material. Treat as informed opinion, not settled fact.
• Below 40% — Mostly model knowledge. The answer rests almost entirely on what the underlying language model already knows. Useful as a starting point; verify before quoting.

How DNA45 calculates it

For every claim in an answer, the system looks outward for support. Five signals compose the final percentage.

1. How many independent sources back the claim. One source supporting a fact is informative. Five independent sources supporting it is decisive. The score grows with the count, with diminishing returns.
2. How credible each source is. Not every URL on the internet is the same. NASA, Nature, the Lancet, peer-reviewed journals, government statistics agencies, established universities, Reuters, AP — these carry weight. Anonymous blogs, content farms, and untraceable forums carry far less.
3. How recently the source was published. Facts age. A 2025 measurement on climate, a 2026 court ruling, the latest population census — recent material outweighs older material when the topic is one that changes.
4. Whether the sources agree with each other. Consensus across independent sources sharply raises the score. Visible disagreement lowers it, regardless of how prestigious any single source is.
5. How topically relevant each source is. A paper that mentions a topic in passing counts less than a paper that is centrally about it. The system measures topical match, not just keyword presence.

Why we use external sources to back facts

This is the part that matters most.

The language models that power DNA45 are extraordinary. They carry an enormous, structured map of human knowledge — how concepts relate, how arguments compose, how language behaves. We take pride in that map. It is not a small thing.

But the map is frozen at the moment of training. The world is not. Facts change. New papers are published. Court rulings overturn precedent. Measurements get revised. A drug is approved, a planet is reclassified, a record is broken, a leader is elected, a number is corrected. None of that retroactively edits the model.

Mechanisms that connect the model to live information streams can improve recency, but recency alone is not evidence. A claim that appeared yesterday is not inherently stronger than one that has been examined, challenged, and confirmed over time.

The Confidence Score does not reward newness. It rewards validated agreement across credible sources. In many cases, the highest scores emerge from facts that have been repeatedly tested and independently confirmed. In other cases, especially for emerging topics, recent sources may carry more weight, but only when they meet the same standards of credibility, relevance, and agreement.

Recent sources can also meaningfully raise a score when they introduce new, credible evidence that updates, refines, or overturns prior understanding, but only once that evidence is visible and supported beyond a single point of publication.

In short: the system values evidence that holds, not simply evidence that is new.

So a model’s answer is a guide to the shape of knowledge — the contours, the relationships, the vocabulary. To turn that guide into a fact you can stand on, the system has to walk outside, knock on the doors of the institutions where knowledge is born every day — the laboratories, the journals, the academies, the agencies — and ask: does the world today still agree?

That walk is what the Confidence Score measures.

What we mean by “credible source”

The system gives more weight to material that has been through some form of public, accountable scrutiny. In rough order:

• Peer-reviewed journals (Nature, Science, the Lancet, JAMA, IEEE, ACM, etc.).
• Major scientific agencies (NASA, ESA, WHO, IPCC, CDC, ECMWF, NOAA).
• Statistical authorities (Stats SA, Eurostat, US Census, World Bank, IMF, OECD).
• Established academic institutions (universities, research institutes, dot-edu / dot-ac domains).
• Major newswires for current-events claims (Reuters, AP, AFP, BBC, Bloomberg).
• Public-record archives (court filings, parliamentary records, treaty databases).
• Open knowledge bases (Wikipedia, when the article itself is well-cited — the underlying citations carry the weight, not the article).

Sources that are anonymous, untraceable, contradicted by stronger sources, or financially incentivised to make the claim are weighted down or dropped.

What the score is not

• It is not a moral judgment. A 50% score does not mean “DNA45 thinks this is a bad answer”. It means “the external evidence right now is partial”.
• It is not a guarantee. Even a 95% score does not exempt you from your own critical reading. It is a signal, not a verdict.
• It is not static. The same question asked tomorrow can score differently because new evidence has arrived, or because consensus has shifted.
• It is not absolute across topics. A confidence of 70% on a recent geopolitical event and 70% on a basic chemistry fact are not the same kind of 70%. The bands above are guides, not rigid rules.

Why we publish it openly

Most knowledge systems hide their uncertainty. They return an answer in the same authoritative tone whether it is well-grounded or improvised. We think that is a quiet form of dishonesty.

By printing the Confidence Score on the face of every card, DNA45 makes the system’s relationship with truth visible to the user, in real time, on every interaction. You always know how much weight to put on what you just read.

The model gives knowledge its shape. The world gives knowledge its weight. The Confidence Score is what we earn when those two agree.

Where the score appears

The Confidence Score is shown on:

• Every answer card returned by the Shepherd.
• Every research paper card surfaced through the Research module.
• Shared cards — the score travels with the link, so anyone receiving the share sees the same honesty.
• Broadcast sessions — every listener sees the same score in real time as the master reads through.

Related

For how data flows through DNA45, see the Data Policy. For the broader operating philosophy of the platform, see What is DNA45?. For privacy and account controls, see the Privacy Policy.