Method & limits

Everything the site shows can be reproduced from what is written here. The thresholds are in the code, the calibration set is published, and every correction we make is public. The full write-up — the method, the evaluation protocol, the results and their limits — is in the repository (docs/paper.md).

What the numbers are

Six factors, each 0–100, each with the exact phrases that produced it:

Clickbait — the headline hides the payload behind a vague word (“the secret”, “the reason”) or asks a question it never answers.
FUD — fear, uncertainty and doubt wording.
FOMO — urgency and scarcity wording (“only today”, “don't miss it”).
Violence, personal attack and profanity — explicit words, matched against a published list.

The factors are counted separately, on purpose: a straight news report about a real crisis is frightening (FUD) without being a curiosity gap (clickbait). The headline number is the strongest of the six — we never average them into something comfortable.

They also fall into two different families, and the reports keep them apart: the pressure (clickbait, FUD, FOMO — the manipulation language, which is what this tool is about) and the content (violence, personal attacks, profanity — what the story is about). A war headline is high on the content and quiet on the pressure. That difference is the interesting finding, so we do not average the two families into one number.

And the honest limit: these numbers describe the language, not the intent. A headline can be alarming because the story is alarming, and a publisher can use these words without any intention to manipulate. We measure pressure, not guilt.

How a number is made

The page is fetched once and the article is isolated — the navigation, the ads and the “related headlines” blocks are stripped, so a menu word never becomes a score.
The headline and the body are compared against a published list of manipulation phrases using a small multilingual sentence model. No LLM, no black box, no per-article tuning.
Each factor is mapped onto 0–100 with its own ruler (the band below), so the numbers are comparable between factors and reproducible.
Explicit word lists decide violence, personal attacks and profanity; the language rules decide clickbait. A factor with no evidence scores zero — we do not let a background similarity look like a finding.

The thresholds

Low	0–33
Moderate	34–66
High	67–100

The calibration is measured on a published Hebrew headline set. The known limits: sarcasm, quoting and reported speech can shift a score; the evidence panel always shows why a number came out the way it did; and the headline set the rulers were fitted on is a development set — a held-out set is being labelled, and until it exists we quote the test numbers only, never the development ones.

When we are wrong

Every number can be disputed, and readers can dispute it from the result page itself. An editor checks each report against the evidence; the confirmed ones are published in the corrections log with the publisher's answer when there is one, and they become permanent test rows so the same mistake does not come back.

Privacy: the article text is fetched once, scored, and only the scores and a short excerpt are cached. We never profile readers, and the reporters in the corrections log stay anonymous.