Reading a chart, predicting an event and controlling a robot require different evidence. This collection helps define those boundaries before a promising representation becomes a product commitment.
The findings below are reported by the researchers in the cited studies. The evaluation questions are practitioner interpretation, not measured Hyperion client outcomes.
LLaTiSA
LLaTiSA represents time series with both plots and indexed numeric tables, using a staged curriculum and the HiTSR dataset. Its reported work concentrates on reading values, perceiving patterns and semantic reasoning. Predictive inference appears in the broader taxonomy, but the paper does not validate causal inference or operational root-cause analysis. Research source.
Evaluation question: Separate extracting a value, describing a pattern, forecasting and identifying a cause in the acceptance tests.
UniT
UniT learns discrete intent representations through cross-reconstruction across embodiments. The authors evaluate transfer, data efficiency and human-to-humanoid conditioning in selected simulation and robot settings. Generalization on those tasks is not proof of adaptation to arbitrary factory layouts or a substitute for physical-system acceptance testing. Research source.
Evaluation question: Evaluate the target embodiment and task distribution, including situations absent from training and transfer demonstrations.
WorldMark
WorldMark compares world models through adapters that translate common movement controls and measure response, motion, memory and visual quality. The revised study includes 10 models and 500 cases. Its scope excludes object dynamics and event-based interaction, so movement-control performance alone does not establish a complete simulation capability. Research source.
Evaluation question: List the interactions the product requires and identify which ones the benchmark does not exercise.
OpenMobile
OpenMobile combines an environment-memory graph with learner and expert policies to gather recovery trajectories for mobile agents. The paper reports results on mobile benchmarks, notably AndroidWorld. Those results do not establish iOS support or general coverage of every application; environment diversity and reinforcement-learning gains remain limited. Research source.
Evaluation question: Test recovery from wrong screens and interrupted actions on the exact operating system and application versions.
COSPLAY
COSPLAY separates skill selection from a second agent that mines rollouts and refines descriptions of skill effects. Its experiments concern game tasks. These effect contracts describe observed state changes; they are not independently validated safety guarantees and do not establish suitability for robots, vehicles or healthcare systems. Research source.
Evaluation question: Check whether a skill's stated preconditions and effects remain valid when the environment changes.
Define the capability precisely
Replace broad requirements such as reasoning or autonomy with observable tasks. Specify the input, acceptable output, operating conditions and failure response, then identify which parts of that contract the research actually exercises.
