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

The story · Chapter 3

One post, from a phone to a verdict

A real fake from the August floods, followed through every step. Every later chapter looks closer at one of them.

The post is a TikTok video. A muddy torrent tears through a hillside town, buildings folding into the water. On-screen text claims hundreds dead and more than a thousand missing. It is being shared as the Bhotekoshi flood. It is not: Newschecker Nepal traced it on 28 August 2026 to the flash flood at Dharali, in India's Uttarakhand, on 5 August 2025.

A reader sends a linkSeen this link, post or file before?yesAnswer at oncenoDownload, take keyframesRead the postthe claim, places, numbersMatch the footageagainst footage of known originGather evidencenews, fact-checks, the webDraft the verdictEnglish and Nepali, with sourcesFixed rules sort the caseA checker decides and publishesThe reader is toldThe footage is remembered
One post through Hamsa. The dashed line is the memory: what a checker publishes is matched the next time it appears.
  1. Second 0A reader asks. Someone unsure forwards the link to Hamsa's Telegram bot. It replies at once with a report number and a page that shows how the check is going.
  2. Before any AIIs it a repeat? If this exact link, post or file was reported before, the report joins the earlier one, and a published verdict is sent straight back.
  3. SecondsThe video is fetched, and up to six keyframes are taken at scene changes: enough to recognise the footage, cheap enough to show a model.
  4. About 15–30 sThe post is read. A vision model reads the caption and frames and writes down the claim, the places, any date, the numbers, the on-screen text and any watermarks.
  5. About 1–2 sThe footage is compared with footage of known origin. The match is strong: the same shot as Dharali, proven point by point.
  6. SecondsEvidence is gathered from news, fact-checks and the web, and a reasoning model drafts the verdict in English and Nepali with its sources.
  7. InstantThe rules sort it. False, footage from elsewhere, nothing conflicting: a quick check, at the top of the desk.
  8. When a checker looksA person decides. The post's frame sits beside the Dharali original. The checker reads, maybe edits a sentence, and publishes. The reader is told on Telegram.
  9. Next weekIt comes back as a Facebook reel: a different file on a different platform. Its frames match the footage the checker just published, so it arrives ready for one click.