V MAX

Why it is different

the one rule

The agent never invents a timecode.

Deterministic signal processing produces a ranked set of legal cut points. The model selects an index. It does not get to name a moment in time.

The evidence

This is not a stylistic preference. At hour scale, every video language model measured scores between 0.001 and 0.115 mIoU on temporal grounding, while a 2021 CLIP baseline beats all of them at 0.269. The paper attributes 85% of failures to search, and finds a retrieve-then-ground hybrid recovers 6.7× — arriving independently at this architecture.

The scope matters, and we would rather state it than have it used against us: on short clips those same models reach around 36 mIoU. The collapse is a long-video problem. That is also the problem worth solving — nobody needs help finding a moment in a twelve-second clip.

So we do not ask the model to find the moment. We ask it to choose between moments signal processing already found.

arXiv:2606.12300

What is genuinely ours

And what is not.

Treating an edit as a document is not new — that is OpenTimelineIO, an Academy Software Foundation standard with a decade behind it. Approval workflows are a funded category of their own. We are not going to claim either. Underneath the one rule, three things are actually ours:

Rational integer time

Time is a pair of integers, never a float, and a float in a time field is rejected at parse with an error explaining why. OpenTimelineIO calls its type RationalTime and stores two doubles. A hundred frames at 29.97 accumulate zero drift here.

Position-independent cache keys

A render chunk is keyed on its own content, never its position on the timeline. Shift a clip without changing it and it is still a cache hit — so a ripple edit, the most common edit there is, stops being the most expensive one.

Reason-carrying operations

Every operation records why it exists. That is what turns a version diff into something a client can actually approve, rather than a list of mutations.

Said plainly

Being agentic is table stakes and someone shipped it before us. The property we have not found anywhere else is the constraint: no other tool stops the model from authoring a timecode. That is the whole argument, and it is the one worth checking.

See the pipelineTry it

Maths finds the moments. The model chooses between them.