A browser extension that reads the piece in front of you and reports what its evidence, reasoning, context and language are doing — then opens every point of that reading to the community, to support or challenge with evidence.
In closed testing. The Chrome Web Store listing comes next.
Your disagreement is the product.Everything you grade or comment on is used to improve the grader.Why ›
A grader that nobody argues with is a grader nobody is checking. So when you think a grade is wrong, Shared Signal asks you to say which observation is wrong and attach something — a source, a correction, a reason. That is not a complaints box. Those grades and comments are the material we measure the engine against and tune it with, and they are the only thing that makes it get better at this.
The engine is measured against the judgements, never against who made them. Whether your name shows on the page is your choice, made fresh each time you post. We do not sell personal information and we do not run advertising. In aggregated or de-identified form, with no name attached, contributions may also be used in other products as Shared Signal develops — the privacy policy says how.
Graded A or B. Sourced, reasoned, and in context — not a finding that any of it is true.
Held up worst
Graded D or F. Thin sourcing, weak reasoning, or missing context — not a finding that any of it is false.
Everything people have run
One card per article, not per person — if four people run the same story, that is one entry with a count of four. Nothing here says who ran anything. Pages that live behind a login — intranets, shared documents, email — never appear at all; subscription articles are graded for the person who ran them and are not published either.
How the grades work
Four dimensions, each graded A to F independently, each a judgement about the text rather than about the publication.
Evidence & Sourcing
How many sources, how independent, and how proportionate to the weight of the claim. Is anything attributed, or asserted bare?
Logic & Reasoning
Does the conclusion follow from what is presented? Judged as given, without assuming the premises are true or false.
Context
Is the background a reader needs in order to weigh this present, or missing?
Tone & Bias
Is the language built to inform, or to provoke?
Factual Accuracy is not graded, and that is deliberate
During calibration, the model graded a real, documented assassination as a fabricated premise at 95% confidence — because the event happened after its training data ended. It did not say it was unsure. It said false.
The other four dimensions are judgements about the text and need no knowledge of the world, which is why they hold up. Checking whether a claim matches reality is a different job, and it needs to search current sources rather than answer from memory. Until it can, this dimension reports nothing and counts for nothing. The claims are extracted and kept, ready for when they can be.
A grade here describes one article. It is never an average for a publication, and it is never a statement that something is true or false. If you think a grade is wrong, the reading's Community Review lets you support or challenge the specific observation you disagree with, with evidence — and that applies to publishers as much as to readers.
