CoulResearch lab
Log inStart free

01Coul research · model card

Virality is not magic.It's measurable.

Coul reads the opening seconds, pacing, retention risk, visual hook, and audience context before a post goes live.

Explore the methodAnalyze my content
78Viral score

02Training recipe

We train forpatterns, not personalities.

The model learns relationships between creative structure and observed outcomes, then validates them on creator-separated examples.

Skateboarder filmed in warm evening light
00:01.2
Creator recording a short-form video
00:03.3
Food creator preparing a recipe
00:05.4
Technology creator demonstrating a product
00:07.5
Short-form creator video example
00:09.6
Short-form video performance example
00:11.7
  1. 01
    Media representation

    Frames, audio, transcript, timing

  2. 02
    Outcome pairing

    Performance signals normalized by context

  3. 03
    Creator split

    Held-out creators reduce memorization

  4. 04
    Score calibration

    Raw outputs become useful score bands

03Signal anatomy

Four layers of signal.

Skateboarder in motion used to illustrate frame-level signal analysis
Beat 01 · 0:00–0:03

Motion hook detected

Subject reads in frame one

Risk at 0:06.8

sun hitsconcretebreathedrop incommitweight overpopfloatland cleanroll away

Coul turns a reel into evidence you can edit.

04Inference path

From frames to a forecast.

Coul doesn't look for one magic feature. It combines hook strength, pacing, quality, retention risks, transcript, and visual-hook notes into a structured report.

Inspect the output schema
  1. 01Source

    Owned video or supported public reference

  2. 02Parse

    Frames · audio · transcript · timing

  3. 03Score

    Hook · pacing · quality · risk

  4. 04Calibrate

    Context-aware 0–100 estimate

  5. 05Explain

    Beat breakdown · edits · prompts

Skateboard video source0:16
Frames
Audio waveform
Transcript

Sun hits concrete. Deep breath. Commit. Pop. Float. Roll away.

Timing
Virality estimate
78Viral score
0100

Move the payoff 1.4s earlier

High priority
Explanation
Beat breakdown0:00 Hook (sun hits)0:02 Set up (concrete)0:04 Commit0:06 Pop0:12 Land clean
Edit suggestions

Move payoff earlier. Tighten beat two.

Visual-hook notes

Lead with motion + light.

Reusable prompts

Write three hook variants for this edit.

05Evaluation & limits

A score needs context.

Virality is a moving target. Platform shifts, audience history, account size, timing, and novelty can all change the outcome.

Illustrative calibration plotPredicted score on the horizontal axis and observed outcome on the vertical axis, with an uncertainty band around the calibration line. The annotations distinguish well-calibrated, under-confident, and over-confident regions. An orange outlier marks a viral result caused by an outside platform event.

Illustrative geometry · audited values publish with each model card

05Evaluation protocol

How we validate a score.

Each model version is tested on creator-separated examples, calibrated against observed outcomes, and published with its evaluation date and known limits.

  1. Creator holdoutExamples from the same creator never appear across both training and evaluation.
  2. Context normalizationPerformance signals are compared within relevant audience and platform contexts.
  3. Calibration auditPredicted score bands are checked against observed outcomes and uncertainty.
  4. Versioned reportingMeaningful model changes trigger a new evaluation date and model card.
AI content intelligence

Start creating with clarity, not guesswork.

Predict virality, generate content ideas, build workflows, analyze performance, and schedule across every channel from one intelligent workspace.

Create accountBook a demo
No payment at signupExplore before checkoutPaid terms shown before payment