Watch closely,
trace the claim.看清画面，
追到证据。

The video makes its case quickly. Within the first forty-five seconds it says that the world’s population is growing exponentially, that it will pass ten billion this century, and that the land used for livestock is larger than Africa. Each comes with a confident number, and a viewer has no time to check one before the next arrives. In other videos the claim sits only on screen, in a chart or a slide, which is why FactLens also reads the picture.这条视频的论证推进得很快。前四十五秒里，它说世界人口正在指数级增长，本世纪会突破一百亿，畜牧业用地比整个非洲还大。每一句都带着笃定的数字，观众还来不及核对一句，下一句就已经到了。在另一些视频里，说法只出现在画面上的图表或幻灯片中，所以 FactLens 也要读画面。

FactLens takes the video as a case file. It does not decide whether the video is “good”. It finds every statement that can be checked, checks each one on its own, and shows where each verdict came from.FactLens 把视频当作一份案卷来处理。它不评判这条视频「好不好」，只做三件事：找出所有可以核查的陈述，逐条单独核查，再把每个结论的依据摆出来。

An investigation starts by collecting everything before judging anything. FactLens gathers three kinds of evidence, each from a different layer of the video.调查从收集证据开始，先不下判断。FactLens 收集三类证据，各自来自视频的不同层面。

1. A

   Testimony: what was said口供：说了什么

   The platform’s subtitles if there are any; otherwise a local Whisper transcription, with every line tied to a timestamp.优先使用平台字幕；没有字幕时，用本地 Whisper 转写，每一句都对应一个时间点。

2. B

   Scene photographs: what was shown现场照片：画面上有什么

   ffmpeg pulls key frames at scene cuts, or at fixed intervals when the picture barely changes; in this case, 12 frames. A vision model reads the text, formulas and chart values on each one.ffmpeg 按场景切换抽取关键帧，画面变化很少时改为定时采样，这一案共抽取 12 帧。视觉模型逐帧读出上面的文字、公式和图表数值。

3. C

   The record: what is known档案：已知的事实

   For each claim, a live web search, keeping the pages the verdict relies on.针对每一条主张联网检索，并保留结论所依据的网页。

A sentence often carries more than one claim. The first sentence of this video is a good example, and it shows how FactLens reasons.一句话里常常不止一个主张。这条视频的第一句就是例子，也最能说明 FactLens 如何推理。

Each verdict is reached in two steps. First a model with live web search researches the claim and writes down what it found. Then a second pass turns that research into a fixed form: supported, refuted or uncertain, with the reasoning, the sources and a confidence the model assigns itself. Keeping the two steps apart lets the web search and the fixed output format work without getting in each other’s way.每个结论分两步得出。第一步，带联网搜索的模型调查这条主张，写下查到的内容；第二步，再把调查结果整理成固定格式：支持、反驳或不确定，附上理由、来源和模型给自己的置信度。两步分开，联网检索和固定格式的输出就不会互相干扰。

All five claims, pinned to one board. Pick a card to read its case notes.五条主张，钉在同一块推理板上。点选卡片，查看每一条的结案记录。