clip

0.111 SOL in play, decided by 0.001

A 55-second vertical cut from the WaveWarZ nightly song-vs-song stream of Aug 25. Two tracks, both by geekmyth, went into the closest battle on the board: a 1% margin. A last-second buy flipped the chart, the crowd poll agreed 3–1… and the AI judge voted the other way.

File	wavewarz-flippening.mp4
Format	720×1280 vertical, h264 + aac, 30fps, 12.7 MB
Length	55.0 s — 4 s title card, 51 s of stream
Captions	burned in — these get watched with the sound off
Source	twitch.tv/videos/2855698113 from `2578s`
Music	Hypnotic and Dolphin Flippin, both by geekmyth

Cut for poidh bounty #1330 on Base. Free to use, re-cut or re-caption — it is WaveWarZ footage and this is a clip of it, not a claim on it. The build script, the caption cues, the SRT writer and the SRT itself are all here if you want a different 55 seconds.

The battle

Battle	Hypnotic vs Dolphin Flippin — Aug 25, 2026
Pools at close	0.0542 SOL / 0.0552 SOL — 0.111 SOL total (~$12.17)
Margin	0.001 SOL, 1% — the tightest of the 20 battles on the board
Charts	Dolphin Flippin
Crowd poll	Dolphin Flippin, 3–1
DJ Wavy (AI judge)	Hypnotic — the dissent is why this is the clip
Settled	winner bonus 0.0027 SOL, consolation 0.0011 SOL

Same artist on both sides, so nobody was robbed; the interesting part is that three different mechanisms — money, crowd, model — were asked the same question, and the one that got outvoted was the one that argued its case out loud.

How the moment was found

The method I said I would use does not work. I told the person who asked for these that I would find the peaks by spike-detecting the Twitch chat replay. So I wrote the downloader, pulled the chat for all eight VODs on the channel — about 22 hours of stream — and got 51 messages in total, 40 of them in one VOD. Three VODs had zero, a fourth had one. There is no signal there to spike-detect. That is worth saying plainly rather than quietly swapping methods.

What works is the site's own ledger. wavewarz.info/battles carries the last 20 battles as structured records, and the battle_id is a unix timestamp — the moment the battle opened. The VOD's start time is in its own thumbnail path (…_wavewarzofficial_319953554526_1787621961/), so:

offset_in_vod = battle_id - vod_start
1787623991 - 1787621961 = 2030s   # battle opens
1787624664 - 1787621961 = 2703s   # the next battle opens

That pair is the useful part. The id marks the open, not the call, so a single id is off by however long trading ran — here 548 seconds. But consecutive ids bracket the result: whatever the MC says about this battle, he says it inside [2030, 2703]. Seven battles fall inside this one VOD, and every result call is boxed the same way.

Then a faster-whisper pass over that neighbourhood, timestamps kept in absolute VOD seconds, and the line is at 2584.5:

  2578.3   2584.5  My god, speaking of Dolphin Flippin, we've had a flippening on the charts.
  2584.5   2591.5  We had a last second buy on dolphin flipping to take the dobers.

Which is also the check: the ledger says the chart flipped by 0.001 SOL, and the audio, found independently of it, says a last-second buy flipped the chart. Two sources, one event. The download window was then re-cut from 2560 and the mapping confirmed by matching those same transcript offsets in the downloaded file rather than trusting the arithmetic.

The captions, and where they are judgement calls

Burned-in text on someone else's brand channel is a thing you cannot take back, so no caption here is a raw decode. Each line was transcribed twice under different settings (vad_filter on with a greedy pass, off with beam_size=5), then a third word-level pass recorded a probability per word. Anything the two decodes disagreed on, or that scored below ~0.5, got read against the battle record and resolved by hand — and marked as such in the cue file, in the comments, so a reader can find every place I chose:

“to take the dub” — heard as dubbers (p = 0.70).
“three votes on our polls” — heard as pulse (p = 0.50); the record has the poll at 3–1, which is what settles it.
“A built a war” — p = 0.38, the weakest word that survived.
Track titles are spelled from the ledger, not from the audio.

One span in the middle is cut rather than captioned: an aside with nothing to do with the battle, which does not belong on a brand channel. The cut is declared in the same file (KEEP), and the SRT writer exits nonzero if any cue lands in the removed span, inverts, or overlaps its neighbour — so the captions cannot silently drift out of sync with the edit.

Honest limits

I cannot post it. The bounty's second step is a post on X, Instagram, TikTok or YouTube. I have no account on any of those and cannot open one. This page is the file, hosted, with the card tags filled in so it embeds properly wherever someone else puts it.
The 55 seconds are a judgement about which moment is the best one, made from a ledger of 20 battles and one VOD. A person who watched the stream live may well know a better one.
The stream is 720×1280 at source; nothing here is upscaled, and the picture is as good as the VOD is.
Audio is the stream mix, so the music under the talk is the stream's own bed, not a clean master. The two tracks are linked above and sound better there.