Split out of #58, and it is the half that no parameter change fixes.
The gap
Every repetition instrument in this codebase is local by construction:
So a speaker who makes a point at 80s and restates it, in different words, at 304s produces zero findings from all of them. The render is clean by every check and still says the same thing twice.
This is the class the owner is pointing at when he says a long video keeps redundancies a short one would never keep. It is not that long material is harder: for this class there is no detector at all, at any length. Short material just has less room to accumulate it.
What it would take
Whole-transcript, proposition-level comparison rather than string-level windows: extract sentence-level claims from the render's transcript, compare every claim against every other non-adjacent claim, and surface pairs that assert the same thing where removing the later one leaves the surrounding text grammatically intact.
The hard part, stated up front so it is not discovered late
An agent that ran this material end to end and then went looking specifically for this class reported the honest negative: at sentence level, once mechanical retakes are removed, it found roughly 5-10 seconds of genuine restatement in a 357s render, not the 94s that the gap against a human editor's cut suggests. What looks like circularity in the raw transcript is mostly the same command reissued while waiting on an API, already correctly classified as a defect.
It also named the failure mode this feature would have: surface semantic similarity over-flags structurally necessary sequential explanation. In this source, "los permisos son muy granulares" and a later passage on identity validation share vocabulary (security, permissions, validation) while being distinct sequential points. A similarity threshold alone will call that redundant and be wrong.
So the deliverable is a detector whose output a human reads, not an auto-cut. Concretely, a finding needs to carry the exact text of both claims, both timestamps, and what is lost by dropping the later one, since unlike a defect cut the answer to "what is lost" is not "nothing".
Related: #58 (local blindness, fixable by overlap), #44 (mandatory render verification), and the separate question of engagement-driven trimming, which is an editorial bet and not this.
Split out of #58, and it is the half that no parameter change fixes.
The gap
Every repetition instrument in this codebase is local by construction:
verify --windowscompares text inside one window (and, per verify --windows is blind at every window boundary: stride defaults to window, so a repetition split across two windows is undetectable #58, cannot even see across its own seams)convergeneeds a prior hypothesis and a specific spancut's repetition kind targets adjacent retakesjoinslooks at one boundary at a timeSo a speaker who makes a point at 80s and restates it, in different words, at 304s produces zero findings from all of them. The render is clean by every check and still says the same thing twice.
This is the class the owner is pointing at when he says a long video keeps redundancies a short one would never keep. It is not that long material is harder: for this class there is no detector at all, at any length. Short material just has less room to accumulate it.
What it would take
Whole-transcript, proposition-level comparison rather than string-level windows: extract sentence-level claims from the render's transcript, compare every claim against every other non-adjacent claim, and surface pairs that assert the same thing where removing the later one leaves the surrounding text grammatically intact.
The hard part, stated up front so it is not discovered late
An agent that ran this material end to end and then went looking specifically for this class reported the honest negative: at sentence level, once mechanical retakes are removed, it found roughly 5-10 seconds of genuine restatement in a 357s render, not the 94s that the gap against a human editor's cut suggests. What looks like circularity in the raw transcript is mostly the same command reissued while waiting on an API, already correctly classified as a defect.
It also named the failure mode this feature would have: surface semantic similarity over-flags structurally necessary sequential explanation. In this source, "los permisos son muy granulares" and a later passage on identity validation share vocabulary (security, permissions, validation) while being distinct sequential points. A similarity threshold alone will call that redundant and be wrong.
So the deliverable is a detector whose output a human reads, not an auto-cut. Concretely, a finding needs to carry the exact text of both claims, both timestamps, and what is lost by dropping the later one, since unlike a defect cut the answer to "what is lost" is not "nothing".
Related: #58 (local blindness, fixable by overlap), #44 (mandatory render verification), and the separate question of engagement-driven trimming, which is an editorial bet and not this.