Can I compare scores between two different content types?

No. A score summarises the findings a particular rule set produced on a particular input. A 92 on marketing copy and a 92 on application code mean different things and should never be ranked against each other.

Is a high score a guarantee?

No. It means the rules that ran found nothing above their tolerance. Whether the argument holds, or whether the decision behind the code is right, remains a human judgement.

Why does the score fall less for the fifth finding than the first?

Deductions decay by a factor of 0.72 per finding, in severity order. One systemic problem takes a large bite; a cluster of small ones still adds up without reducing everything to zero.

Do old reports change when the rules do?

No. A saved report keeps the findings and score it was produced with, and the changelog records which version of the rules was live on that date.

Methodology

How a score is built, and what it cannot tell you.

A number nobody can reconstruct is not evidence. These are the weights, the decay and the bands, written so you can disagree with a specific part of them.

Calibrated once, trusted for the whole run.

What is measured

Four deterministic checkers read one input and return findings, metrics and a count of the rules that ran. Nothing is inferred by a model, nothing leaves the browser, and the same input produces the same report whenever it is run.

Metrics are measurements, not faults: line counts, reading ease, bytes per pixel, checksum presence. They appear beside the findings so a reviewer can see the evidence rather than only a verdict.

How the score is calculated

The score starts at 100. Each finding deducts its severity weight, multiplied by a decay factor of 0.72 raised to the finding's position when sorted by severity. The result is rounded and floored at zero.

The decay exists for one reason: without it, a file with six medium findings scores the same as one with a leaked key. With it, the most serious finding takes the largest bite, and a long tail of small ones still registers.

Critical−22
Something that must not ship: a credential in source, unparseable data, an image that cannot be decoded.
High−12
A defect with a realistic path to harm or visible failure: an execution sink, a swallowed error, a broken outline.
Medium−6
Quality and consistency problems that erode trust or cost later work: placeholders, dead links, flat rhythm.
Low−2.5
Measured observations rather than faults: line length, adverb density, duplication.

Bands and verdicts

The grade and the plain-language verdict are both derived from the score. They are deliberately conservative: the top band is described as ready to publish rather than good.

90–100APublish-ready
80–89BShips with small fixes
70–79CFix before publishing
60–69DFix before publishing
55–59DNot ready to ship
0–54FNot ready to ship

The limits, stated plainly

It does not detect who or what wrote the work. Authorship detection is unreliable and is not offered.
It does not judge whether a claim is true in the world, only whether it is sourced and phrased responsibly.
It does not run your code. Static reading finds structure and patterns, not behaviour.
Image findings come from the decoded pixels and embedded metadata of the file you supply.
Reports are capped at the 24 highest-severity findings, and text input is read to 200,000 characters.

Why human review is a different instrument

A reviewer is not a better checker. A reviewer answers questions the rules cannot: whether the argument holds, whether the audience is right, whether a likeness is acceptable. Their output is a decision with a stated scope and a name, and it costs hours rather than milliseconds.

The two are meant to run in sequence. Automated findings clear the mechanical questions so the review starts at the part that needs a person.

Questions this raises

Can I compare scores between two different content types?: No. A score summarises the findings a particular rule set produced on a particular input. A 92 on marketing copy and a 92 on application code mean different things and should never be ranked against each other.
Is a high score a guarantee?: No. It means the rules that ran found nothing above their tolerance. Whether the argument holds, or whether the decision behind the code is right, remains a human judgement.
Why does the score fall less for the fifth finding than the first?: Deductions decay by a factor of 0.72 per finding, in severity order. One systemic problem takes a large bite; a cluster of small ones still adds up without reducing everything to zero.
Do old reports change when the rules do?: No. A saved report keeps the findings and score it was produced with, and the changelog records which version of the rules was live on that date.

See it applied

One file, one score, every finding named.

The report shows the findings behind the number, so the arithmetic is checkable rather than trusted.

Run a check Request a specialist