The Hamsa Book · contents

The story

  1. 1The flood, and the videos that came after
  2. 2A triage desk, not an oracle
  3. 3One post, from a phone to a verdict

How it works

  1. 4Seeing footage before: three fingerprints
  2. 5Finding the evidence, in two languages
  3. 6Models propose, rules decide
  4. 7A memory, not a library
  5. 8Under the hood: one laptop, open models

The proof

  1. 9The numbers, and what went wrong
  2. 10What Hamsa does not do yet
← Back to Hamsa

How it works · Chapter 5

Finding the evidence, in two languages

Where the sources come from, and why a source that merely mentions the same flood is not enough.

A verdict is only as good as its sources. Hamsa gathers evidence from four places for every post, and keeps up to eight passages, at most two from any one article, so one long report cannot crowd out the rest.

  • Trusted Nepali news and official sources, collected every 15 minutes by a news watcher.
  • Fact-checks published abroad, through Google's ClaimReview index, because AFP, BOOM and Alt News often debunk the same clips that circulate in Nepal.
  • The open web, through a search engine Hamsa runs itself (SearXNG), for events the trusted outlets did not cover.
  • What the platform shows on the post: a Community Note on X, or TikTok's "AI-generated" label.

Relevant is not the same as related

Early on, one fact-check roundup about Nepal's floods was cited against nearly every flood post, because its words matched. But a fact-check about a different video of the same flood proves nothing about this one. So before a fact-check counts, a model reads it beside the post and answers one question: is this about the same footage? If it cannot tell, the fact-check is dropped.

Then a reasoning model (gpt-oss, open-weight) reads the claim, the footage match and the passages, and drafts the verdict: what it found about the footage, the event and the figures, a label, a confidence, and the passages it relied on. A second check makes sure the Nepali says what the English says.