View A. Every extracted claim plotted by confidence (x) and salience (y), coloured by tradition. The whole method lives in the split between the two: a wild claim can be low-confidence and high-salience and is kept and watched (top-left), never deleted. Click a point to flag it (watch → bad extraction → clear), then Download patch to export your annotations as JSON you can commit.
Point opacity encodes the epistemic rating (metaphysical → mythic → interpretive → descriptive): fainter = more abstract. Quadrant split at 0.5.