World models and embodiment

A world model is a learned, action-conditioned generative model of an environment, used as a simulator: roll a policy forward inside it, explore counterfactuals, re-plan when reality diverges. The motivation is rollout throughput; the usability bar is fidelity — the deployed policy should not be able to tell the generated world from the real one.

Embodied agents hit constraints software agents share but rarely name. Vision models are bad at comparison — detecting a small change between two states is exactly the capability review depends on. Named gaps move and unnamed ones do not. And social transparency about intent and provenance is a capability, not a finish.

Verify rows execute on merged main against real CI, run by someone other than the implementer — closer to a real-world rollout than to a simulation, and a deliberate position rather than missing tooling. Cheap simulated verification would only ever be worth its correlation with reality, and the stated fidelity bar is the standard any replayable-eval ambition here would have to meet.

Who raised an item, what work produced a brief, which role acted — answerable from the artifact without an agent narrating it. That is the same property the robot demonstrates, reached by a different route, and it is why the desks can run silent. The perception finding is the sharper transfer: present old and new state together in one context rather than two sequential reads — a diff-based review does this; read-the-file-twice does not. Numeric verification before visual judgment is already the rule, now with an external reason attached.

Digital cousins: generate variations of a known failure scenario to harden a guard against the neighbourhood of a fault rather than the single instance. Designed nowhere, recorded here as the transferable idea from this theme.

8 of 11 talks shown — the ones that reach the most concepts in this theme. Every sighting, per talk, is on the concept pages above.