Keep a memory of the world.

The seed world map has three observations from three distinct named sources. Two entries contain nonblank causal link text and one does not. At eighty pixels per entry total observations, linked and unlinked measure two hundred forty, one hundred sixty and eighty. These sentences describe a changed light and moved object, but the application never reads a camera or performs model inference. A recorded link is a user claim about causality, not a demonstrated intervention, timestamped experiment or evidence of a real world model. The default entries assign High, Medium and Low confidence, one each. At one hundred pixels per entry the three bars are equal. Filtering High shows one card while total observations remain three and the two causal link labels remain in the full stored data. The labels have no probability calibration or numeric confidence formula. A filter changes visibility rather than deleting observations or validating their interpretation. Export includes every observation even when a filter hides some, so the visible card count and exported row count are deliberately different. Adding one observation from an existing source with a blank link makes four observations while unique sources remain three and linked entries two. At eighty pixels per count observations, sources and linked entries measure three hundred twenty, two hundred forty and one hundred sixty. Blank observation or source prevents the add. The browser stores data locally. On first load data references the seed array directly, so adding also mutates that reset seed: the native reset still shows four after this addition. A later fresh load with stored data can behave differently. The actual JSON download includes observations, causal labels and limits of no live camera access or model inference. Storage and export preserve authored claims, not prove them true.