Content DNA
Content DNA · — YouTube videos · — genres

Every genre edits to its own beat. Ads cut —× faster than cooking.

Median cuts per minute, by genre

One bar per genre · — videos measured frame by frame
Grounded
—
of the evidence behind its advice checks out against the measurements
Specific
—–—
a blind judge preferred its critique over a generic summary
Not causal
Patterns, not proof
nothing here shows a style raises views; links to engagement are correlations
1 · Problemand the instrument

Analytics say where viewers left. Not what the edit was doing.

Existing tools score, summarise or list shots. None explains construction with timestamped fixes.

ToolMeasures structureExplains constructionTimestamped fixes

Measure. Interpret. Check everything.

  1. Measurecuts · loudness
  2. Listentimed words
  3. Watchwhat's on screen
  4. Fingerprint30 features
  5. Critique3–6 timed fixes
  6. Verifyevery citation
2 · One videosecond by second

Every observer on one clock.

Story beats, screen content, cuts, speech and loudness share one timeline. Pins are fixes; click one to jump there.

3 · Patternsacross — fingerprints

The first 15 seconds: ad vs. cooking.

One frame every 2.5 s.

Tech reviews talk — words a minute. Vlogs, —.

Each dot is a video; the white tick is the genre median.

Explainers speak at once. Vlogs wait — s.

Seconds until the first voice.

Each genre has a visual diet.

Share of runtime by what's on screen.

Presenter time tracks likes. Probably the audience, not the edit.

Only 1 of — features survives correction (ρ —), and it shrinks to — once audience size is controlled.

Correlation with like rate, within genre, with 95% interval. The 8 strongest features shown; highlighted = survives correction.

4 · Trusttesting the observers

Every observer failed once. Cross-checks caught each one.

Cut detector

A morph is not a cut.

— precision · — recall
Vision model

It misread the clock.

— → — boundaries on a real cut
Transcription

It heard speech in music.

— invented transcripts caught
Critique

It quoted lines nobody said.

— of cited evidence verified

Measuring makes advice checkable.

Evidence verified, by kind

With measurements vs. transcript only (— videos).

Blind judge scores

Preferred —–—. A model judging a model is weak evidence.

—
fix moments recur on a re-run
—
story-beat labels stay the same
—
tag wording overlaps: words drift, moments hold
5 · Limitswhat we know, and what we don't

Feedback with an address. Not yet proof it works.

✓ Observed

  • Genres differ sharply in pacing, speech and format
  • Measurements make advice checkable
  • Cross-checks caught every observer's failure

! Untested

  • No A/B test: no fix is shown to raise views
  • — videos from search, not representative
  • Like rate is a weak proxy; no retention data
—
API cost per video
—
timestamped recommendations
— h
of video measured
All — videos
Cost breakdown
StepCallsUSD

Whole project including evaluation and re-runs: —. Estimated from logged tokens.

Methods

Shots: PySceneDetect AdaptiveDetector. Motion and colour: OpenCV at 2 fps. Loudness: librosa RMS. Transcript: whisper-1 word timestamps, screened for invented passages. Screen formats and calls to action: gemini-3.5-flash on the YouTube URL. Critique: claude-sonnet-5 with structured output; every cited metric, quote and visual is re-checked by a deterministic validator. Judge: claude-opus-4-8, blind. Statistics: Kruskal-Wallis with Benjamini-Hochberg, Spearman with bootstrap intervals and within-genre permutation tests. Regenerate everything with analysis/*.py.

Content DNA Analyzer · research prototype. Every number on this page is read from the pipeline's result files (results/story/). Recommendations are hypotheses to test, not performance predictions.