World inspector

Every world is generated from a seed: a causal DAG drawn from an anonymised template, or up to eight templates bridged into a chain (the depth control). Left column: what the agent can know. Right column: what only analysis and humans can. The same boundary, shown at every level of representation.

seed compound depth 0 = atomic · n = n bridged templates chained complex variables (z = a + j·b) — off by default

Agent surface

Observer truth

Prose

Diagram

The same quantities, deliberately without edges: the structure on the right is exactly what the agent does not get.

Diagram

JSON

Variable atlas

One card per variable. The shape does the identification: a bell is a gaussian source, a straight line a linear response, a power curve a multiplicative one, a blow-up an exponential, an S a saturating response. Curves are drawn from each variable's actual parameters, and each family shares one scale across its cards — a narrow sigma is a tall spike, a steep slope is visibly steeper — so cards compare against each other, not only against their own box. Multi-parent mechanisms run the curve along one parent: the strongest for additive families, and for multiplicative the exponent farthest from 1, since x^0.95 would draw as a straight line and defeat recognition by silhouette.

Truth is public on purpose: any reader of the public repo can compute it from the seed, so secrecy is not what protects the benchmark. The boundary is process-level — nothing on the agent's path calls the truth endpoint — and anonymisation guards against retrieval from training, not against a person reading the answer. ← demo · the evidence