VTCF

Bangla clickbait agent

VTCF · AUTONOMOUS VIDEO AGENT

An agent that watches the video before you do.

VTCF reads the title, samples the frames, and calls out Bangla YouTube clickbait when the footage won't deliver what the headline promises.

vtcf-agent

IDLE

$paste a Bangla YouTube URL — I'll watch it before you do.

Case files

Open a case from the 805-video test set

Pre-computed results load instantly. Hard cases show where BanglaBERT title-only fails but full VTCF gets it right.

Why VTCF

Titles can lie.
Frames can't hide.

Text-only classifiers miss clickbait when the headline reads like legitimate news. VTCF pairs language understanding with visual evidence, so the verdict reflects what the video actually shows — not just what it claims.

  • Reads the title's promise with a BanglaBERT text encoder
  • Watches hook, context and delivery frames with a Vision Transformer
  • Cross-attends text against visuals to catch the bait-and-switch
8,047

Videos studied

Human-labeled Bangla YouTube videos in the research corpus

805

Test videos

Held-out 10% evaluation set for the primary VTCF model

Research findings

What the hard cases taught us

Across the 805-video test set, 29 text-blind hard cases were evaluated. The data tells the story better than we can.

Temporal divergence

Genuine videos move. Clickbait stalls.

Genuine · mean TDS 0.644Clickbait · mean TDS 0.383
0.20.40.6HookContextDelivery0.6440.383

Mean temporal divergence accumulated across the video timeline. Endpoint means are measured study values; the curve shape between checkpoints is illustrative.

Headline metrics

Accurate on the easy cases. Unbeaten on the hard ones.

99.63%

F1 score

Full VTCF model on the main test set

100%

Hard-case rescue

All 29 text-blind hard cases recovered by visual evidence

Finding 1

Visual-Temporal Contradiction

Clickbait changes its visuals far less than genuine content — bait recycles static footage instead of telling a story. That gap (TDS 0.383 vs 0.644) is the detection signal.

Finding 2

Production-Style Deception

Polished clickbait fools every text channel at once: plausible speech, clean on-screen text. When all the words lie consistently, only the frames can disagree.

The rescue

29 text-blind cases, 29 recovered

All 29 testable hard cases that defeated the text-only channel were correctly recovered when visual evidence joined the decision.
The science

Powered by VTCF : Visual-Temporal Contradiction Framework

Distinctive finding

Clickbait videos show less visual change on average than genuine videos because clickbait often reuses static footage rather than bait-and-switching content.

Score guide

What is Temporal Divergence Score (TDS)?

A compact signal for how much a video's visuals move across three checkpoints — always shown next to the frames and explanation in your results.

Explore a sample score

0.52/ 1.0

Moderate shift

0 — steady1 — strong shift

Noticeable visual change across the three sampled moments in the timeline.

1Hook
2Context
3Delivery

VTCF samples these three moments to summarize visual change over time.

Temporal Divergence Score (TDS) is a number from 0 to 1 that summarizes how much a video's visuals change across three moments : hook, context, and delivery. It gives you a quick read on temporal visual change, not a verdict by itself.