Independent field notes / 0106

Research lab

Eight editorial observation exercises from the supplied field school. These are prompts to investigate, not reports of hands-on testing.

Separate demo from product

A lab video proves a direction, not latency, cost, availability or creator control at scale.

Write what was shown, what was measured and what remains unknown.

Weights, data, license, limitations and evaluation details matter more than a launch headline.

Check permitted use and missing evaluations before reuse.

Generative visuals must still reflect the deterministic world state players are acting inside.

Identify which layer owns collision, inventory and damage.

Cloud generation, edge inference and local rendering each spend part of the response window.

Mark where one dropped frame becomes a broken interaction.

A world model must keep objects, materials and motion coherent from frame to frame.

Look for flicker, identity drift and geometry that changes under motion.

Generation is useful when artists can constrain, edit, version and reproduce results.

Ask whether the same prompt can be directed and revised predictably.

Classifiers can assist review but still produce misses, false positives and context errors.

Document escalation and appeal paths alongside model accuracy.

Graphics progress matters only if creators can ship it and players across devices can use it.

Compare fidelity gains against hardware, cost and bandwidth requirements.