An AI that cannot make a number up.
Ask Sigil about anything on the page and it answers from the same read the page rendered with, citing what it used. This is not a policy we ask a model to follow. It is a property of how the answer is built.
The mechanism
Fabrication is not caught afterwards. It is impossible to express.
Every grounded value exists only as a fragment minted server-side from an actual read. The model is handed opaque references and may cite them, never author them. A number it writes as prose has no fragment behind it, so it cannot wear a grounded badge: there is nothing to cite.
Guidance wears a dashed outline, so the distinction survives a grayscale render and does not depend on colour alone.
Never more confident in chat
If the page will not conclude, neither will Sigil.
A fragment carries the same confidence state the page renders. Where the sample sits below the floor, the answer says what is missing rather than rounding up to a conclusion the pixels refused to make. A null is an honest none, never a zero.
Two verbs
Sigil answers in the drawer and drafts in your inputs.
The same character, doing two different jobs with two different promises. An answer cites read-model values it can prove. A draft is a proposal against a field contract and cites nothing, so it is marked as a suggestion and nothing changes until you accept it.
Show me
Every claim links back to the surface it came from.
A cited answer carries the read it used, the rule version behind it and the sample floor it cleared. Follow it and you land on the surface that produced it, so the answer is checkable rather than merely confident.
An assistant that would rather say it does not know.
Every module is part of one platform. You get all of it, and the tiers differ by scale rather than by which capabilities are switched on.