How this reads a text
Clef marks the devices, and spots sources and causes left unanswered. Clairic names who is speaking. An LLM picks the words, writes the plain lines and the claim, and looks for conflicts.
Pressure counts the writer's own voice and any citing of an authority. Quotes and reported speech are left out. It counts nine kinds of sign: the five marked in the text and four more right often enough to count.
How it argues: an LLM notes where the writer stands, then Clef is asked of each sentence, one part at a time, thirteen parts in all. Conclusions and reasons are read from all the answers together by a small fitted step. The margin beside each sentence names its parts, and a dot marks one read with none found. A mark or part whose answer is within .10 of its line is drawn lighter. These nine parts agreed with Claude's labels on opinion pieces. The lines above them count the sentences in two parts at once. Other kinds of text are untested.
How it's built: for each reason, Clef is asked which nearby claim it supports, told where the writer stands. A reason is placed only where that answer is at least .50 sure. On 30 opinion pieces, placed reasons matched Claude's labels about 7 times in 10.
Says why someone acts: a sentence giving an intention or hidden reason for what people do, marked only where Clef is very sure. On three sets of opinion pieces, about 9 in 10 such marks matched Claude's labels, though most such sentences go unmarked.
Not read: whether the facts are true, what was left out, what the writer intends.
Marks are noisy. On labelled propaganda, about 1 in 5 matched a human label.