Perceptron Mk1.5 links media observations to proposed function calls

A timestamp or marked region gives a reviewer a place in the original recording to check the model’s account.

The distinction between observation and action is the important part for builders. A timestamp or bounding region can help a person inspect what the model appears to have noticed. A function call then proposes an operation for the surrounding application to execute. Neither is evidence that the scene was understood correctly, and the model does not itself determine which external actions should be allowed.

The current model card gives concrete operating constraints. It lists a shared context of 36,864 tokens and a maximum output of 8,192 tokens; audio consumes the same context budget as other material. Soundtracks in video require explicit enabling. Responses can contain text and structured content, but the model does not generate audio.

The quickstart provides a Python SDK and a first authenticated request. Published input and output prices are $0.15 and $1.50 per million tokens respectively. Those rates are useful starting points for a trial budget, not a cost per minute of video or a guarantee of application-level affordability.

For a first evaluation, choose a narrow recorded task with known answers: locate an event, recover its timestamp and ask the model to propose a harmless next step. Compare each output with the source recording. A successful demonstration is worth pursuing when its errors, latency and total cost are measured as carefully as its fluent explanation.