Your calls are not τ-Voice
τ-Voice and the October 2026 Grok, GPT-Live and Gemini boards are still public exams. The only eval that answers whether your agent works is a held-out set of your calls and your kitchens.
Isaac Sim is not a damp sponge
NVIDIA Isaac Sim, Isaac Lab and Cosmos can scale lighting, layouts and object variation. They do not give you a damp sponge or a sticking drawer. Those still have to be collected wet.
A creator app is not a brief
Paying people by the minute to film chores is a capture funnel, not a written skill. Hour-counts without reject codes, task-not-completed, camera-on-ceiling and an eval split by person are not a dataset.
A GPAI training summary is not a dataset card
The EU GPAI public training-content summary is mandatory for models on the Union market. It is a model-level page about sources and processing, not a dataset card for the human set you add.
Speech preference is not task success
October 2026 voice boards show Gemini winning Speech Agent Arena preference while losing on Tau Voice agent tasks. Preference data and task-success eval are different objects. Do not use the same batch as both.
Paying a creator is not consent for the rest of the house
Figure Index pays creators to film chores. Sensory data covers the creator and the environment. Paying that person is not bystander consent and not a commercial training pack you can buy later.
You cannot evaluate a Live agent on isolated utterances
Full-Duplex-Bench-v3 scores Live voice agents on real human disfluencies and chained tool use. Isolated utterances are the wrong exam for a product that has to book, interrupt, and self-correct.
Figure’s compute is not the skill
Figure’s Nscale deal buys up to 100k GPUs for Helix. Adcock still names data and compute as the constraints. Buying GPUs does not buy the skill, the kitchen or the rights.
Voice eval is not a transcript
OpenAI’s GPT-Live eval guide says a transcript misses interruption, overlap and tool use while audio continues. WER on a booth clip is the wrong exam.