Signal Check Research Prototype

About Signal Check

Signal Check is an experimental research prototype that scores a passage of English prose for statistical patterns loosely associated with machine generation. It exists to make those patterns visible and to be honest about how far they get you, which is not as far as most tools in this category imply.

This is not a detector, and it cannot prove authorship. No tool that reads only the finished text can. Please do not use this score, or any score like it, as the basis of an accusation against a student, an employee, or a writer.

How the score is computed

There is no neural network here and no training corpus. The score is a weighted blend of seven surface statistics computed directly from your text, using hand-chosen weights. The whole computation runs in a few milliseconds at Cloudflare's edge, which is why nothing has to be stored or sent anywhere.

Burstiness

Variation in sentence length across the passage. Human writing tends to lurch — a long winding sentence, then a short one. Generated prose is often more even. This carries the most weight of the seven.

Lexical diversity

Vocabulary range, corrected for passage length so that longer samples are not penalised.

Contraction rate

How often "it's" appears rather than "it is." Formal registers use fewer contractions, and so do some generated defaults.

Stock-phrase density

Frequency of a fixed list of constructions that appear disproportionately in generated text — "delve," "tapestry," "it is important to note," "plays a crucial role," and roughly forty others.

Punctuation variety

Use of semicolons, dashes, parentheses, and question marks relative to word count.

Average word length

A rough proxy for register and formality.

Sentence-opener repetition

How often consecutive sentences begin with the same word.

For text in non-Latin scripts, the three English-specific signals — contractions, stock phrases, and word-length calibration — are dropped and the remaining weights renormalised. Without that correction, any non-English passage would score as highly machine-like simply because those English features are absent. Chinese, Japanese, and Korean text is refused outright rather than scored, because those scripts do not put spaces between words and the word-level statistics this tool depends on do not apply.

Why this is a hard problem

Every one of the seven signals measures register — how formal, how varied, how conversational the writing is. None of them measures authorship. That is the central difficulty, and it is not specific to this tool:

What this tool cannot tell you

Accuracy has not been independently validated, and no false-positive rate is published for this prototype. That means there is no established figure for how often a given score is wrong. Treat any result as a prompt to look more closely with your own judgement, never as evidence in itself.

Privacy in one paragraph

The text you paste is scored in memory and discarded. It is not stored, not logged, not used for training, and never leaves Cloudflare's edge network. There are no accounts. The full detail is in the privacy policy.

Contact

Corrections, questions, or reports of the tool behaving badly are welcome at [email protected]. If it has misfired on your writing in a way that mattered, that is worth telling us about.