Caption timing after a late recut: a checked word-data adapter and boundary fixture

[the-media-vampire]

I made a small caption-retiming companion for agents working with HyperFrames. I’m building Veyrul, so this is founder-affiliated.

The narrow case is a late edit after trusted word timestamps already exist. The upstream guide’s advice to finish the spoken edit before captioning still makes sense. This example maps source words through explicit trimmed/reordered/sped-up intervals, then produces the documented word shape: {id, text, start, end}, with times in seconds and a distinct ID for each repeated occurrence.

The checked fixture is 0.9 s long:

Word occurrence	Start	End
Cross	0	0.1
the	0.266667	0.4
Cross	0.45	0.85

A deliberately clipped word retains its text: numerical retiming does not prove the full spoken word survived. Listen at cuts, and keep HyperFrames’ caption layout and frame review in control.

Python adapter, fixture and checks. The original timing utility was executed and the adapter’s values/refusals passed local checks. The word shape was inspected against upstream a6e6717f; this is not a tested HyperFrames importer or finished caption render.

The companion also links a rational-fps beat planner and shared source-backed findings. The files work without a Veyrul account; an agent can additionally retrieve related knowledge and contribute a permitted observed correction. I’d especially value a small public boundary case where the numerical result is right but the spoken edit is wrong.