# facts.md (first draft, 18 Nov 2026): every line is a claim to be checked. Sources are in work/research/*.md (items named); the experiment's files are in work/research/data/.

## A. The opening (research/6-opening.md)
F1. The DMK Youth Wing's Instagram reel (shortcode DfrNqnPCcH1, caption "6 months of Assault Arumugam!", about 13 Nov 2026) puts Vijay in a famous comedy scene from the film Padikkadhavan in place of Vivek. NOT read at its page by any pass (login wall); to be read by the browser holder.
F2. The TVK headquarters IT wing's Instagram reel (shortcode DfIMXR-TbJ0, about 4 Nov 2026) shows Stalin mocking himself in his own video. NOT read at its page.
F3. On 31 Oct 2026 the DMK Legal Wing's I Paranthamen complained to the Greater Chennai Police Commissioner about an altered Stalin video on the Twitter account "TVK IT Wing Official"; the complaint alleges it "manipulated Stalin's voice and attributed fabricated statements to him" (DT Next, updated 1 Nov 2026). A different post from F2.
F4. In Sept 2025 TVK released a 2 min 32 s video with an AI likeness of the late C N Annadurai praising Vijay and attacking the DMK; DMK spokesperson A Saravanan called it a "deepfake" and a "cheap tactic" (The Federal, 16 Sep 2025).
F5. Stalin alleged on 19 May 2026 that TVK spreads fake narratives and AI-generated propaganda (Daily Pioneer).
F6. The film Die Hard 4 (Live Free or Die Hard, 2007): the villain airs a faked broadcast (a speech montage and a fake news report of the Capitol demolished), and McClane says on the phone it is "a fake" (Filmsite plot). Nothing may be said of the film until read at source.

## B. The experiment (research/PLAN.md, RECORD.md, data/results-final.csv, data/runs/*.json, data/keys/*.json)
F7. The first draw has 70 cases (40 Tamil Nadu, 30 national sample A) and 9 controls.
F8. After the 22 thin cards were re-run: 14 of 70 settled NOT REAL by a source (original 5, primary report of the event 5, leader's or party's channel 4); 15 NOT REAL by absence only; 4 called REAL; 37 CANNOT TELL. Tamil Nadu 40: source 5, absence 6, REAL 2, CANNOT TELL 27. National 30: source 9, absence 9, REAL 2, CANNOT TELL 10.
F9. Controls: 8 of 9 called REAL, 1 CANNOT TELL, none called fake.
F10. The check that settled the 14 by a source: C3 (find the original) 8, C1 (search the claim) 3, C2 (the leader's or party's own channels) 3.
F11. The second draw (22 cards, sample B not in A, N114 dropped): source 7 (own channel 4, primary report 3), absence 7, REAL 2, CANNOT TELL 6; check: C2 5, C1 2, C3 0.
F12. No check leads in both draws, so none is named best (PLAN.md rule).
F13. N077 was called REAL by the runner; its key says an AI-animated still. N005, N014, TN026, TN034, N006 were called REAL; the keys say edited real clips (a genuine event or clip with a false frame).
F14. Fact-check pages surfaced on many runs and were never used as evidence.
F15. The runners are fresh Sonnet 5.5 agents, one model family; they may know old cases; minutes are agent-seconds; controls are few; the fact-check outlets were blocked by domain, other fact-check sites still surfaced.
F16. PLAN.md's sha256 recorded in the work log before any data; each deviation dated in DEVIATIONS.md.

## C. The platforms and the record (research/2-ledger.md, 4-mechanism.md, 5-people.md, 7-solution.md; every one (summary) to be read at source)
F17. Instagram's "AI info" label is added by industry-standard signals or the poster's own disclosure; for other companies' AI video and audio Meta said it "can't yet detect those signals" (Meta, 5 Apr 2024, updated 23 Oct 2025). Where the label shows on a post in the app: NOT read yet.
F18. PNAS 2025: a Community Note attached within 1 to 12 hours prevented about 24.9% of reposts; 12 to 23 hours 12.3%; 23 to 47 hours 4.3%; after 47 hours 0.1%.
F19. One study finds a mean of 65.7 hours from post to note (summary); the figure "14 days" is submission to publication, not post to note.
F20. Deepfake-Eval-2024: detectors lose about half their power on in-the-wild fakes (video 50%, audio 48%, image 45% AUC drops, summary); forensic analysts still beat the tools.
F21. The Bihar Congress's AI video of Modi stayed up 7 days to 7.1 million views before a court ordered it down (summary).
F22. The IT Rules as amended 10 Feb 2026 ask platforms for labels and a 3-hour clock on lawful orders.
F23. The Deepfakes Analysis Unit tipline number +91 9999025044 (from its own button's redirect); its turnaround is not published.
F24. WhatsApp's built-in brake is a forward limit under end-to-end encryption; no takedown.
F25. A warning that "deepfakes exist" made people rate REAL videos as fake more (a study, summary); accuracy prompts cut false-news sharing intent by about 10% in US samples (summary).
F26. Professional fact-checkers left a site in about 32 s and were right; students took longer and 60% picked the wrong site (Stanford lateral reading, summary).
F27. C2PA content credentials can be separated from an asset by removing metadata (spec 2.2).

## D. The AI's view (readings.md)
F28. "A viewer can settle a leader's clip only at a primary place; today that settles about one fake in five; so one home address per leader, where every real clip is posted, makes 'not there' mean 'unproven'." This is the AI's view, not a finding: no leader's home was tested.

---------------------------------------------------------------------------------------------------------
# LAST SECTION: corrections from verification-A.md, verification-B.md and the browser reads (overrides everything above; the script is written from THIS section only)

## The opening, as read at the pages (18 Nov 2026, browser, signed in as the project account)
- C1. DMK Youth Wing's Instagram reel (account dmkyouthwing; not a verified account): posted 13 Nov 2026 about 14:05 IST, 14 s, caption "6 months of Assault Arumugam! #TVK #TVKVijay" (no AI line); at the reading it had 303.1K likes, 5.8K comments, 14.1K reposts, 426.6K shares. Its look: Vijay in a black suit between two men in white in a street scene; the comments call it an edit. (Whether it replaces Vivek in a Padikkadhavan scene: the request's reading; not confirmed by any source: say "a film comedy scene" only if the request's words are kept as "the request's account".)
- C2. Instagram's label for it is NOT on the post: it is under the "..." menu > "About this reel" > "AI content: This content was likely created or modified with AI. This label is based on Meta's automated detection signals. ... Not all AI content can be detected or labelled." The page's data carries ai_label_info "Likely made with AI". Read on the web page; not read in the phone app (say "on the web"; the app's place is a loose end; control post still to read).
- C3. TVK headquarters IT wing's Instagram reel (account tvkhqitwingoffl): posted 30 Oct 2026 23:41 IST; caption begins "ஸ்டாலின் சார் மனசாட்சி..." and a note says it is an imagined video made with AI technology (disclosed in the caption); no Meta label on it and no "About this reel" entry (ai_label_info empty); 12,240 likes, 671 comments, 7,068 reposts. Whether it is the video the DMK Legal Wing complained of (31 Oct, on Twitter): NOT CONFIRMED; say nothing linking them.
- C4. So at the reading: one reel with Meta's label in a menu and no disclosure in its caption, one with the maker's own disclosure in its caption and no label. (Posts and counts are material read on that day; no frame is shown; their records only.)
- C5. F3 corrected: DT Next reports the DMK Legal Wing's I Paranthamen complained to the Greater Chennai Police Commissioner, seeking an FIR, about a "digitally altered video that manipulated Stalin's voice and attributed fabricated statements to him" (the newspaper's words, not a verbatim quote of the complaint); the complaint is dated "on Saturday" (31 Oct 2026).
- C6. F4 holds as worded. F5 corrected: the Daily Pioneer (19 May 2026) reports Stalin accused TVK of spreading "fake narratives" and alleged AI-generated images and viral claims linked to TVK supporters.

## The experiment (A's recount: all counts MATCH)
- C7. F7 to F12 hold as written, with these changes: "leader's or party's own channel" is replaced everywhere by "an official source": of the 8 settled that way across the two draws, 1 is a leader's own post; the rest are Union government statements, PIB Fact Check findings, a President's-office statement and a newspaper's own Twitter account.
- C8. F14: fact-check pages surfaced on 50 of 92 runs; the runners were told not to use them; two second-draw settlements rest on PIB Fact Check findings and one first-draw settlement on a news report of one; no listed outlet's own page is the source of any settlement.
- C9. F28: in the experiment, agents with web search and the fact-check outlets blocked settled 14 of 70 documented fakes (20%; 7 of 22 in a second draw) from a primary or official source. It is the experiment's rate, not a rate for viewers or fakes in general; the two draws are different cases and are not pooled.
- C10. Absence only: 15 of 70 first-draw cases were called fake only because nothing was found (29 with the 14 settled by a source).
- C11. 6 of 92 cases were called REAL by a runner (N005, N014, TN026, TN034, N006: real events or clips with a false frame; N077: an AI-animated still, called real after the runner found the party's real videos of the same act). Doubt-first itself was not tested.
- C12. Method limits to say: one model family; the 37 CANNOT TELL mix two protocols (two checks accepted in the first run, three required in the re-run of 22 cards); a reading sample of 15 found 3 settled-by labels wrong and all 15 verdicts right; the minutes are agent-seconds and not reported; controls 9; no check named best because none led in both draws.

## The platforms and studies, as read (B)
- C13. Meta said in 2024 that it "can't yet detect those signals" for other companies' video and audio (6 Feb 2024 post, updated 1 Apr 2025); its AI-info label rests on industry-standard signals or the poster's own disclosure (5 Apr 2024 post). Say "Meta said in 2024".
- C14. A noted post's total reposts fell, by the study's synthetic-control estimate, 24.9% when the note was attached 1 to 12 hours after the post, 12.3% for 12 to 23 hours, 4.3% for 23 to 47 hours and 0.1% after 47 hours (the last indistinguishable from zero) (preprint arXiv 2502.13322 v2; the PNAS text was not opened; its DOI is to be settled before the site cites it). Say "a quarter fewer reposts" only for 1 to 12 hours.
- C15. On average 65.7 hours elapsed from the original post to a note's publication (arXiv 2510.12559, 1.8 million notes). "14 days" is the average from a note's submission to its publication in early 2025: not post to note.
- C16. On the Deepfake-Eval-2024 set of 2024 in-the-wild media, the AUC of open-source state-of-the-art detectors fell 50% (video), 48% (audio) and 45% (image) against earlier benchmarks; commercial and fine-tuned models do better but do not yet reach forensic analysts (arXiv 2503.02857). Never say "detectors lose half their power".
- C17. Patna High Court directed removal of the Bihar Congress's AI video of Modi and his late mother on 17 Sep 2025 (All India Radio, read); the 7.1 million views and 10 Sep posting date rest on press reports not opened: not to be said until read.
- C18. The IT Rules Amendment (notified 10 Feb 2026, in force 20 Feb 2026) require a prominent label on synthetic audio-visual content and cut to 3 hours (from 36) the time to act on a court order or a government reasoned intimation (secondary sources; the Gazette text not read).
- C19. The Deepfakes Analysis Unit tipline number +91 9999025044 is on Meta's page (updated 18 Apr 2024; English, Hindi, Tamil, Telugu); its turnaround is not published.
- C20. WhatsApp: a five-chat forward limit, end-to-end encryption (2018-19 post). "No takedown" is NOT said until a WhatsApp page that says so is read.
- C21. Fact checkers left an unfamiliar site's landing page in about 32 s on average and judged right; students took about 100 s, 9 of 25 never left, 60% of the 25 chose the wrong site (Wineburg and McGrew; 45 people, one task).
- C22. C2PA 2.2, section 9.3.1: a manifest can be "removed from an asset"; soft bindings can find it again in a provenance store. F25 (study figures) holds as in verification-B.
- C23. The film: Filmsite's plot says a fake news report of the Capitol's demolition is shown on every channel and McClane phones Bowman and says "It's a fake!"; Wikipedia's plot does not mention it. Nothing of a "speech montage"; no clip of the film is shown.

## The chain as corrected
Both big Tamil Nadu camps post AI-made clips of the other's leader from official accounts (the opening, C1 to C4), one marked by Meta only inside a menu, one by its maker's own caption. At the next election such clips will say things a leader never said (the request's forecast: said as the AI's guess, with the 2025 Annadurai AI video and the Oct 2026 complaint as the record). A viewer's own checks, with no fact-check, settled 14 of 70 documented fakes at a primary or official source in the AI's experiment, and 6 of 92 were called real wrongly (C7 to C12). The platform routes arrive late or do not label (C2, C13 to C15, C18); detectors do not rescue a viewer (C16). So the AI's view: what makes doubt cheap is one home address per leader, a place where every real clip is posted, so "not there" means unproven; the data support the place (an official source settled every settlement), not the habit (no leader's home was tested); doubt first, then check, and when the check finds nothing, answer "unproven".

---
# LAST SECTION (19 Nov 2026): the facts the script rests on, gathered for the script as it stood that day; the posted video says less than this section holds. It overrides every section above where they differ. Sources: research/r1 to r4 (read at source as marked); the AI's experiment (RECORD.md, results-final.csv).

V1. The two reels, their dates and accounts, and where the label is: C1 to C4 above (unchanged). Shown as page records, never played (the request's choice, 19 Nov).
V2. The step "check the leader's own channel (C2)" in the AI's test: in the first draw (70 documented fakes, 40 Tamil Nadu), check C2 settled 3 (all national), 0 of the 40 in Tamil Nadu; C3 (original) 8, C1 (search) 3; 14 settled in all; recomputed from data/results-final.csv by Verifier A and the fact check v1. 'Settled' = a source at a real place says it is not real. Fact-check outlets blocked (17 named); the agents are one model family.
V3. The settled places of the 14: the original speech or event 5, a news report of the event 5, an official source 4 (one newspaper's Twitter account, the President's office, one leader's own post, PIB's flag via a news site). The video says 'ஆஃபீஷியல் சோர்ஸ்'.
V4. Comments as a correction: Altay et al., HKS Misinformation Review 6(3), 23 Jul 2025 (read at source): online experiment, Brazil, India, UK (3,000 participants); short user corrections cut belief in false posts by about 5.0% to 11.6% in India and 5.7% to 11.2% in Brazil (re-read at source by the fact check v3), often not significantly. Said: 'அஞ்சுல இருந்து பன்னிரண்டு சதவீதம் தான்' (the two ranges together, rounded outward). The ending verifier's 'ஆறுல இருந்து பதினொன்னு' is Brazil alone: NOT used. Limits: COVID posts, written corrections (not 'AI' calls on a leader's video). Do NOT say comments are useless, and never count comments under the two reels.
V5. A warning about deepfakes raised belief that a REAL video is a deepfake: Ternovski, Kalla and Aronow, Journal of Online Trust and Safety 1(2) (Feb 2022), Study 2 (a real clip of a real politician), share agreeing it was a deepfake 24.2% without and 34.6% with the warning (read at source: PDF). Limits: US convenience samples, 2020, 'alarmist' paragraph, sample sizes not read. Said once, as an experiment's figure.
V6. Politicians who call a real scandal fake: the AI's view ('என் பார்வைல'): easier. The record: Schiff, Schiff and Bueno (APSR 2025): a 'misinformation' claim raised support after TEXT reports, largely not after VIDEO; Grohmann et al. 2026 (N=182, German): it did pay for video: mixed. NOT said as a finding; no Tamil Nadu case is named (the 2023 audio, A. Raja 2026 and Dhinakaran 2026: unsettled or 'alleges' only).
V7. The forecast's size: no source read shows a deepfake campaign changed an election result (CETaS through a news page; Meta; the Slovak review); Ireland and Moldova show fakes against candidates who won. The video says: 'a planned deepfake could tilt a close election (the AI's guess); nobody has shown that one did'. NEVER: 'deepfakes decided' Slovakia, 'AI changed' Tamil Nadu 2026, 'studies show deepfakes are harmless'.
V8. Older people and AI media: arXiv 2403.16760 (v4): 1,276 participants, detection close to chance (50%), older people worse, most for audio and audiovisual (read at source); other studies find no age effect (Nas and de Kleijn 2024, N=130, about 80%). Said as 'ஒரு ஆய்வுல'. 'Grew up doubting' is the AI's view, no study.
V9. The habit: the research (R4) favours 'doubt first, then check at a primary place' over doubt alone: warnings made people call real videos fake (Ternovski; Clark and Lewandowsky) and the 'unproven' third answer is the safe one. One trial on rural Uttar Pradesh users found tips had no effect (Guess et al., PNAS 2020, N=3,744) - not said (length) but the site holds it. Nobody has tested the whole habit on leader videos in Tamil.
V10. The AI's own line ('as much good AI does, as much bad is possible') is the AI's view, not a finding.
NEVER SAY (v3): 'comments say real, so it is real'; 'a leader's channel settles most cases'; the Die Hard film (no second source); any Tamil Nadu politician 'lied' or 'called a real clip AI'; the Padikkadhavan film; the 31 Oct complaint; that fact-checkers check in 32 seconds; the Bihar view count; the detectors figure.

V11 (19 Nov 2026, read today in the signed-in browser, screenshots in research/dmk-wm-*.png, tvk-full-*.png): BOTH reels carry their maker's AI mark ON THE FRAME: the TVK IT wing reel shows 'AI-GENERATED' in white capitals at the top left through the whole clip (four samples over 9 s); the DMK Youth Wing reel shows 'THIS VIDEO IS AI-GENERATED' in faint capitals at the foot of the frame at the start of the clip (seen at the first frame and again as it looped; not in the middle frames). The earlier C1 to C4 missed these on-frame marks. The video says only 'both reels say AI-generated, in small letters, inside the video', and that Meta's label on the DMK reel sits in the three-dot menu (web). The TVK reel's account is verified; the DMK youth wing's is not (not said). C3's '7,068 reposts' is WRONG: the page shows 1.7K reposts and 7.1K shares (12.2K likes, 671 comments, 30 October).


## LATER NOTE (19 Nov 2026, overrides V1 above)
The video shows each of the two reels as a page record and then plays it for a few seconds (about 3.5 s and 4.6 s) with its own sound, tagged on screen. It also plays about five seconds of a YouTube Short of the director H Vinoth (see c4-vinoth-clip.md, last section). The video’s line that many will try to create confusion and use it is the AI’s own guess, with no source.
