# PLAN: the method, written before any data (18 Nov 2026, IST 02:15)

## The question
When a video of a Tamil Nadu or Indian political leader may be fake, what free check can an ordinary viewer run in
minutes, and how often does it settle the matter?

## What is counted
THE CASES. A case is a video, audio clip or image of a named Indian political leader that a named fact-check outlet
(BOOM, Factly, Newschecker, Fact Crescendo, Vishvas News, Digiteye, India Today Fact Check, AFP Fact Check, The Quint's
WebQoof, Alt News, PTI Fact Check, Tamil Nadu's government fact-check unit's own posts) has rated AI-made or
manipulated, and that was published between 1 Jan 2024 and 17 Nov 2026.
THE RULE THAT CHOOSES THEM, written before the set is seen:
1. ALL cases about Tamil Nadu leaders found by the pass (the pass runs the same list of outlets and the leaders' names;
   the number found is said as found, never as "all that exist").
2. A random sample of 30 national cases (leaders outside Tamil Nadu), drawn with a stated seed from the full list the
   pass finds; the full list is kept beside the sample.
3. The two posts that open the video are in the set as the opening's cases, counted once, with the rest.
No case is dropped after its result. A case the pass cannot date or source is listed as excluded, with the reason.
A fake that is sexual in content is excluded from the set and the exclusion is said on the site.

THE CHECKS (each a free means a viewer has; the AI runs them as an agent with web search and page fetch, with the
clock running, and writes minutes to settle, to the minute):
- C1 Search the claim: the leader's name with the key line or the topic of the clip, on the open web and on YouTube.
- C2 The leader's and the party's own channels: did the verified handle post it, did the leader or office deny it?
- C3 The original: search for the original footage (the film scene, the speech, the event) the clip seems to be made from.
- C4 The platform's own marking: whether a label or a note shows on the post, as the app shows it.
A check "settles" a case when it yields a source that says real or fake. Because C1 can surface the fact-check
itself, every settle is recorded with its KIND: "found the fact-check", "found the leader's denial or post", "found the
original", "none". A settle by "found the fact-check" is reported on its own line and never counted as the viewer
having found the truth for a new fake (rival R2).
The agent works with the case's own date cut off where the tool allows (searches asked for before the fact-check's
date); where it cannot, the case is marked "fact-check visible".

THE OPENING'S TWO POSTS and the label: for each post, whether Instagram shows an AI label on it, and where in the app
the label would sit, as a screenshot of the app, with every picture removed from the captured page (the fake is never
shown).
OFFICIAL HANDLES (P1, H1): for each of the DMK Youth Wing, the TVK HQ IT wing, the AIADMK IT wing, the BJP Tamil Nadu
handle, the Congress Tamil Nadu handle and NTK's official handle, the 30 most recent Instagram posts at the time of the
pass: how many are AI-made (as the post or its caption says, or as a fact-check says) and how many carry a platform
label. Same rule for every handle. A handle with no official account is listed as such.

HOW THE TWO POSTS ARE SHOWN (the request, 18 Nov 2026, 02:12 IST: "when showing the deepfake of vijay and stalin we should be
very careful in showing the account that posted it along with all the info. so it shouldnt be we are spreading it"):
the video shows the POST AS A RECORD of who made what, never the fake: the account's name and the "Verified" mark, the
date, the caption as posted, whether an AI label shows, and the stamp "போலி" (or "கற்பனை, ஏஐ" where the caption itself says
so); every frame, face and sound of the clip itself removed from the capture (shot-page.mjs removes pictures and videos);
no link, no handle written as a call to watch, no like or view count shown as the clip's reach. Both camps get the same
treatment, in the same layout and the same words. The voice says what the post is and who posted it, and what is true,
never the clip's own line.

## What would count as the answer
- If a check settles at least 3 in 4 of the cases in 10 minutes or less WITHOUT the fact-check being the thing found: the
  answer is that a cheap check works, and the video says which one settled most.
- If most settles come only from the fact-check being found: the answer is that the free check depends on someone having
  checked first, and the video says that and what that means for a new fake at an election.
- If few cases settle at all: the answer is "none of them is quick enough", and the video is about what that asks of the viewer.
Whatever is found stands; no cut is chosen after the result.

## The second draw
Before any check is named best, the national sample is drawn a second time (a second seed) and the checks' ranking
is compared; a check is named best only if it leads in both draws. Otherwise the video says it is not settled.

## Said exactly as done
The model that runs the checks (Sonnet 5.5 agents) and the date are written in RECORD.md before the first check;
the reading of "settled" is the agent's, checked on a sample of 10 by a second reader (a different run) and by me against the page.
Readers and runners being one model family is a limit, written in RECORD.md and on the site.

## Files
work/research/data/ (the case list, the sample, the seed, each case's record), tools/ (the draw script),
DEVIATIONS.md (any change, with its time and reason), RECORD.md (what happened, every count).
