Research the problem. Prototype the method. Build the system.

In speech, we examine recognition, speaker assignment and dialogue separately. In vision, we examine image and scene interpretation, continuity and the effects of a repair. Each task needs its own source material, baseline and acceptance criteria.

Our initial speech programme compared adapted checkpoints with unchanged foundation models. Broader visual-model and video-production comparisons remain ongoing or proposed work.

The tradeoff: a tightly defined test makes an error easier to explain, but its result may not extend to another language, scene or recording condition.

Prototypes let us examine what happens when speakers overlap, terminology changes, an object loses its shape or an action happens in the wrong order. We retain unsuccessful attempts and inspect the parts that need correction.

Model adaptation, input preparation and workflow changes are different interventions. A selective video repair, for example, should be assessed for both the intended correction and any changes to approved content.

The tradeoff: a targeted correction can preserve more of a sequence, but it may depend on the scene and require manual review. We need comparisons before choosing it for wider use.

Building an application adds work beyond the model: preparing inputs, connecting sources, handling interruptions, presenting revisions and making results available for review.

Turnwise is our AI studio for using and evolving models. Its eight starter templates describe reusable ways to understand needs, assess evidence, plan activities and develop work. Selected speech prototypes are being hardened for pilots. Vision and video work remain research and prototypes, with their own deployment and evaluation questions.

The tradeoff: review and recovery add steps. We assess whether those steps help people find errors and complete the task with less correction.