A real candidate card
They Played Until They Won
8.5“I think losing, brother.”
strong_hook 2:24:47-2:24:51: The speaker argues that losing motivates players more than winning.
Because now it's like they already lost but we're still playing the last game, you know I wonder what motivates them more losing or winning. I think losing brother I think losing remember the one time they lost like 15 games in a row and then they kept they kept playing until they won … We were playing against it was me leery heart I think mbl is there as well And we were taking turns playing through the threes against the Chinese guys And we beat them something like 14 or 15 times in a row and they kept saying 14 go next go next go next until they won … the score in the end was like it was literally like 18 to 1 in the end.
Payoff Cut Short
approvemissing_payoff
missing_payoff 0:28-0:34: The original end cuts off before the answer lands.
Extend the end to include the missing payoff. Window 0:10–0:34 becomes 0:10–0:42; the clip keeps its score and ships.
This is an actual object the finder returned on a real job, not a mock - five reason codes, three timestamped evidence spans, a calibrated 8.5. Scrub the window or pick a code to see the span that backs it. Below it, a reviewer repair: the window was real, the payoff was not finished, so instead of rejecting it extended the end.
Why the gate exists
Uploads are hours long and nobody can watch all of them twice - once to find moments, once to check them. So the job is split between two models with deliberately different jobs. The finder is built for recall: it reads the transcript chunk by chunk and proposes every moment that might make a clip, including quiet, practical, or story-driven ones - its instructions explicitly say not to limit picks to loud viral moments. The reviewer is built for judgment: it sees one candidate at a time and acts as the quality gate, approving, rejecting, or repairing the window.
That split is the whole design. A single model asked to both find and judge would trade recall against taste on every call; separating them means the finder can stay generous and the reviewer can stay strict - and every pick you see carries both sets of writing.
What the finder writes
The finder reads timestamped transcript chunks - lines labelled [M:SS] or [H:MM:SS] - and returns structured candidates, never prose. Each candidate is a card with a fixed shape:
{
"startSeconds": 8687,
"endSeconds": 8732,
"title": "They Played Until They Won",
"hook": "I think losing, brother.",
"interestingnessScore": 8.5,
"candidateType": "story_turn",
"selectionReasonCodes": ["strong_hook", "story_turn", ...],
"selectionEvidence": [
{ "reasonCode": "clear_payoff",
"startSeconds": 8718, "endSeconds": 8732,
"summary": "The final score was about 18 to 1 ..." }
]
}Three rules make the card trustworthy instead of decorative:
- 01Timestamps are absolute seconds, converted by rule. A
[13:29]label becomes809, not1329- and the hook must be the literal first words a viewer hears. Intros, outros, sponsor reads, and housekeeping are excluded by instruction, and nothing under fifteen seconds is eligible. - 02The score is calibrated, not defaulted. Scores are 1-10 with one decimal place, and the rubric bans
7and7.5as lazy defaults - a number has to reflect the specific hook, payoff, standalone context, and rewatch value. Multiple candidates from one chunk must spread by at least 0.2 unless truly tied. - 03Every claim cites its span. Reason codes come from a closed vocabulary - fifteen codes, from
strong_hooktofunny_banter- and each must point at timestamped evidence inside the transcript. A score with no evidence is a vibe; the schema does not allow it.
What the reviewer does
The reviewer sees one candidate at a time: its ID, the allowed start/end ranges, a hook-context excerpt, a full-context excerpt, and any supplied external evidence. It is told to validate that candidate - not to hunt for a different clip - and its answer is equally structured: a decision, defect reason codes, timestamped evidence, and a boundary decision.
The interesting part is the third option between keep and reject. Boundary repair exists because finders get windows wrong in a predictable way: the most common real defect is a window that ends before the payoff finishes. When the moment is real but the end cuts the answer, reaction, or celebration short, the reviewer prefers extend_end with a minimal proposal inside the allowed ranges over throwing the clip away. trim_start, extend_start, and trim_end cover the other three edges; only reject_clip discards.
One thing the reviewer may not do: rescore. Its instructions preserve the finder's number on approved clips - the gate validates quality, timestamps, title, hook, and reason, so the score you see is the finder's calibrated judgment, not a committee average.
What it cannot see
Both models read transcript text, not pixels. That is an honest limitation and the page would be lying without it: the finder cannot tell a face is in frame, and the reviewer is explicitly forbidden from inferring visual facts from the transcript, title, or hook. Codes that would need eyes - visual_reframe_risk, wrong_subject_risk, caption_collision_risk, extraction_unavailable - can only be raised when they cite supplied external evidence by stable ID.
The same discipline runs through the vocabularies: twelve candidate types, fifteen finder reason codes, twenty-four reviewer defect codes, six boundary decisions. Closed lists mean a card cannot describe something the pipeline cannot represent - which is also why every claim on this page maps to a constant in the source.
Signals, not mandates
When viewer-behavior data exists for a chunk, the finder gets it as an ENGAGEMENT SIGNALS section: sections the audience replayed most, speech-rate spikes where the speaker sped up, and chapter labels treated as weak creator-supplied hints. The instructions are explicit that these are supplementary context, never mandates - a most-replayed peak without a transcript-backed hook and payoff is still not a clip, and when a replay peak and a speech-rate spike converge on the same timestamp they can only strengthen a moment the transcript already supports.
Raw-audio signals - energy peaks, silence, confidence anomalies - were removed from the prompt in June 2026 after an internal rewatch-signal study. What is left is the ranking that earns its place: transcript semantics first, viewer evidence as a nudge.
FAQ
Does the AI watch the video?
No. Both models read timestamped transcript excerpts, not pixels. Claims that would need visual evidence - like a reframing or caption-collision risk - are only allowed when they cite supplied external evidence IDs; the reviewer is instructed not to infer visual facts from the transcript.
What is a reason code?
Every pick cites concrete claims from a closed vocabulary - strong_hook, clear_payoff, story_turn, and twelve more on the finder side; twenty-four defect codes on the reviewer side. Each code carries timestamped evidence spans, so a score is never just a vibe.
Can the reviewer change a clip's score?
No. The reviewer is instructed not to rescore approved clips - it preserves the finder's 1-10 calibrated score and only validates quality, timestamps, title, hook, and reason. Its real power is the gate: approve, reject, or repair the window.
What does boundary repair mean?
When the moment is real but the finder's window is off, the reviewer proposes a minimal fix inside the allowed ranges instead of rejecting: trim_start, extend_start, trim_end, or extend_end. The most common real defect is a window that ends before the payoff finishes, so extend_end is the preferred repair.
Do 'most replayed' sections automatically become clips?
No. Viewer signals - most-replayed sections, speech-rate spikes, chapters - are passed to the finder as context, explicitly marked as supplementary and not mandates. A peak without a transcript-backed hook and payoff is still not a clip.
What never makes it into a clip?
The finder is instructed to exclude intros, outros, sponsor reads, and housekeeping, to reject moments that need outside context, to keep every second earning its place with no dead air or filler, and to hold a 15-second minimum. The reviewer rejects late or weak hooks, missing payoffs, and metadata the transcript does not support.
Read the reasons on your own footage.
Upload a long video and every clip it finds comes back with the card - score, reason codes, evidence spans, and the reviewer's call. While the pipeline reads it, the analysis tape keeps the wait watchable - free weekly hours, no card.