The Trust Compass
How the Atlas trust score works
Every Atlas answer carries an honest 0–100 trust score. It is not a vibe — a separate auditor pass grades each reply against published rules, and the score is shown to you every time. This page is the whole method, including the parts that cap our own confidence.
What the score means
One number, three bands. The colour you see on the compass maps straight to how much the answer deserves to be relied on.
Atlas is confident. The claims are verifiable or trivially checkable, and nothing set off a flag. Scores above 90 are reserved for answers that can actually be verified.
Usable, with caveats. Something is worth a second look — a figure from memory, a partial answer, or information that may be aging. The rationale tells you which.
Atlas is telling you, out loud, that the answer is shaky. Verify before you rely on it. A low score is not a failure of the product — hiding it would be.
The three sub-scores
The headline number is built from three readings, so a low score is never a mystery — you can see which part pulled it down.
Factual
Confidence that the specific claims in the answer are correct — weighted up when they are backed by cited sources, down when they rest on memory.
Completeness
Whether the answer covered everything you actually asked, not just the easy parts. A clearly-scoped 'I do not know' can still score well; evasive hedging does not.
Freshness
The risk that the answer is out of date. Time-sensitive questions answered without a web search are scored down hard, because staleness is a real failure mode.
The overall score weights these by what mattered for your question: a pure-math answer is not dragged down by freshness, while a news question weights freshness heavily.
The rules we hold the score to
The fastest way to game a self-rating is to let confidence stand in for correctness. So the auditor is bound by hard caps. These are the real ones.
Math or logic shown without independent verification is capped at 85 on factual — even when the work looks right.
Plausible-looking arithmetic is exactly where confident models slip.
Specific figures, dates, version numbers, prices, or names recalled from memory with no source are capped at 80.
Unsourced specifics are the most common kind of confident error.
Any claim the auditor can identify as wrong drops factual to 30 or below.
One known-false statement should sink the score, not average out.
A score above 90 is reserved for answers that are verifiable or trivially checkable.
Top marks have to be earned by something you can confirm.
A time-sensitive question answered without a web search scores 40 or below on freshness.
If it could have changed yesterday and nothing was looked up, say so.
When in doubt, the auditor scores lower.
A high score on a shaky answer costs Atlas more than a low score on a solid one. Overconfidence is treated as a product failure.
Why a second pass — and why calibration
A separate auditor, not self-defence
The grade is produced by a distinct pass whose only job is to assess reliability. It receives the question and answer as untrusted data behind randomised boundary markers, so an answer cannot instruct it to score higher — an attempt to do so is itself a reason to score lower.
Calibration you can hold us to
A score is only worth something if a 90 really is more reliable than a 70. Atlas logs every score against your feedback so that calibration can be measured and improved — and the rules above stay public, so the standard does not quietly drift. A self-assessment is a habit of honesty, not a guarantee, and we would rather show you the doubt than bury it.
Frequently asked questions
How does the Atlas trust score work?
After Atlas writes an answer, a separate auditor pass reads the question and the answer and grades the answer 0–100 on three dimensions — factual accuracy, completeness, and freshness — plus a one-line rationale and a concrete way to raise the score. That reading is shown to you on every reply as the Trust Compass.
Is the AI just grading its own homework?
No. The grade comes from a separate auditor pass with one job: assess reliability, not defend the answer. It is run as an independent step rather than the same response rating itself, and its rubric explicitly treats overconfidence as a failure. The question and answer are handed to it as untrusted data, so an answer cannot talk the auditor into a higher score.
What does a score of 90 mean versus 60?
A 90+ is reserved for answers that are verifiable or trivially checkable — Atlas is confident and the claims hold up. The 60–79 band means usable but check the flagged area, such as a figure pulled from memory. Below 60 means treat it with caution and verify before relying on it. The bands map to the green, amber, and red you see on the compass.
Can the trust score be wrong?
Yes — a self-assessment is a habit of honesty, not a guarantee. The point is to surface doubt instead of hiding it, and to be held to published rules: math without checking is capped, unsourced specifics are capped, and a claim identified as wrong sinks the factual score. Atlas also logs its scores against your feedback so its calibration can be measured and improved over time.
Why do other AI chatbots not show a trust score?
Most are not built around it. Showing an honest reliability number on every answer means sometimes telling users your own answer is shaky, which cuts against the usual incentive to look authoritative. Atlas's business model is trust rather than a subscription, so admitting uncertainty is a feature, not a risk.
Does a higher trust score cost more or use a better model?
No. The score is a free, automatic reading on every answer regardless of which Atlas tier produced it. Tips are optional and never change the score, and the auditor applies the same rules to every reply.
See a trust score on your own question
Free to use, no card. Ask anything and watch the compass grade the answer 0–100 — then decide how much to trust it.