A look-test for the simulation's renderer, drawn by hand with fake data. The question it exists to answer: can we show belief diverging from truth without it turning into a wall of parentheses?
Figure 1 · primary view
The canvas, and one agent thinking
The bubble is not a new visual language — it is the same renderer at half scale. Because every panel shares one layout computed from ground truth, the mini world needs no labels: position identifies the place. B has never looked at the river, so the river is ghosted; B has inferred fish at the cave, so the fish is amber and carries 🧠 rather than 👁️.
Figure 2 · compare mode
Three heads, one layout
Compare mode is the same view zoomed out with every mind opened at once. Nothing here is a special rendering path — panel two is B's thought bubble, unwrapped. The eye compares three copies of one arrangement, so divergence reads instantly: only the amber fish and the ghosted nodes differ.
Encoding
Three belief states, one hue
agrees
The mind believes it and it is true. Quiet outline, no colour. Most of every panel looks like this.
diverges
Believed and false. The only hot colour on the page, so a glance finds every mistake in a head.
unknown
True, but this mind has no percept of it. Ghosted and dashed — how ignorance is drawn, and B's whole starting state.
deliberating
Teal is machine activity, never belief. A ring while the CEK machine is burning fuel; gone the moment it finishes.
Every believed fact carries a source glyph, which is the provenance system made visible. When B reaches the empty cave, the glyph is what tells it — and the viewer — that this belief was inferred, not seen:
Node positions come from ground truth and every panel reuses them. This is what makes divergence readable at a glance, and it is why the bubble can drop its labels — position already says which place it is.
02
Colour only where a mind is wrong
One accent hue, spent entirely on divergence. Teal is reserved for machine state and means nothing about belief. If a third meaning ever wants a colour, it does not get one.
03
The bubble is the world, recursively
A belief environment is world-shaped, so it renders with the identical code. Nested theory of mind — A's model of B, containing B's own bubble — falls out with no extra work.
04
Only spatial facts are drawn
Horn clauses, goals stacks and continuations are not in the picture. The moment the whole knowledge base gets rendered, the bubble becomes the wall of parentheses this study exists to avoid.
Honest limits
What this does not prove
Hand-placed coordinates, five nodes, two agents, one moment. A real force-directed layout of forty nodes will be uglier, and overflow is unsolved: a node holding six entities has nowhere to put them. Motion is untested — plans advancing and percepts landing are the parts most likely to turn noisy. And nothing here is interactive, so the pan-and-zoom level-of-detail idea, where a bubble expands into a full panel as you approach it, is still a claim rather than a result.
Gate
No visual code gets written until this passes. If the look is wrong, changing it costs an afternoon now and a rewrite later. The substrate work — the graph, the agent loop, perception daemons — is not blocked by this and can proceed either way.