Vijay on your timeline all day without a follow: how?

That is the 2023 code's weight, and it is a weight on the model's predicted chance of a reply, not a count. Twitter's 2026 weights put a reply at about 7 likes (5 against 0.708).

The weights, the like-count pools and the anti-spam rules were first published on 13 Aug 2026, three months after the result; what Twitter ran in March and April 2026 was never published.

In the 2026 code a post is already in the out-of-network index when it is created and at its first like; 32 likes adds it to one more pool.

The comment covers actions taken on a post opened directly, and was written about mass blocking and reporting; the published like-count indexer reads only the count, and the trend list is outside it.

The studies measured liking for unknown faces and shapes, and how true repeated statements seemed; neither looked at media, timelines, parties or votes, and 'rated truer' is not 'believed'.

The 34,000 figure has no named source; the only count from Vijay's side is 1,600 groups with 3 lakh members (August 2023). The funded-IDs charge is one rival-camp operative's claim to one paper, with no evidence shown.

Published code is not the code that ran: the 2023 weights are old, the 2026 weights came after the election, the trend detector was never published, and the weights apply to predicted chances, not counts. You cannot read anyone's reach from a README.

He is the Chief Minister; of course he is in my timeline, he is news. Any leader in power fills the timeline without a fan army.

In the 2026 code a new post is already in the out-of-network index at creation ('post_creation') and at one like ('1fav'); 32 likes adds it to one more pool ('32fav'). github.com

The weights, the like-count pools and the anti-spam rules were first published on 13 Aug 2026, three months after the result; what Twitter ran in March and April 2026 was never published. github.com

The comment covers actions on a post opened directly, written about mass blocking and reporting. The published like-count indexer checks only the count, and the trend list and replies are outside the comment. github.com

Twitter's policy expressly permits 'Authentically coordinating with others to express ideas, viewpoints, support, or opposition towards a cause'; what it forbids is many accounts run to boost a hashtag, inauthentic coordination and bulk replies. help.x.com

In Twitter's 2023 score the reply weight multiplies the predicted chance of any reply and reads no text or tone; the negative weights belong to the viewer's own 'show less often', mute, block and report. github.com

Zajonc (1968) measured how much people liked unknown words, shapes and faces; Hasher et al. (1977) found plausible statements met again were rated more likely true. Neither studied media, timelines, parties or votes, and 'rated truer' is not 'believed'. doi.org

The figure first appeared on TV and in a news site's piece where every line is 'it is said', with no named source. The only count from Vijay's side is 1,600 groups with 3 lakh members (August 2023). deccanherald.com

One rival-camp IT operative's claim to the Times of India, with no evidence shown. timesofindia.indiatimes.com

Meta publishes no weights. Its head says sends per reach matter most and more for unconnected content, but no ratio; figures such as 'a send = 3-5 likes' have no Meta source. transparency.meta.com

SimClusters (the-algorithm at ef4c5eb): a post's embedding is the sum of the interests of the users who like it, each like fading by half every 8 hours (HalfLife = 8.hours); the 'light' index of each community's top posts takes only posts under one hour old (OldestTweetInLightIndexInMillis = 1.hours.inMillis), across about 145,000 communities. This is the reel's 'the first hour matters'. That a crowd liking in the first hour shapes who sees the post next is our inference from this code. Twitter's open-sourced code (the-algorithm, SimClusters), 2023, 2023-03-31

phoenixRankAllCandidateProcessor.strato (commit 3b6919f): favoriteCount >= 32 puts an original post in the '32fav' pool; any liked post goes in '1fav' and every new post in 'post_creation'; retrieval reads '1fav' for 24 h and 48 h and '32fav' for 24 h (phoenix-rankall/src/config/mod.rs). The README says phoenix-rankall is the index that out-of-network retrieval queries. A post is re-indexed at each power of two of its like count (1, 2, 4, 8 ...), at most once a minute, and the indexer reads only the count, not where a like came from. So 32 likes is one more door, not the first. Added on 13 Aug 2026 (commit 47c1bcd), after the 4 May 2026 result: it shows how Twitter works now, not what it ran during the campaign. Twitter's 2026 open-sourced recommendation code (phoenix-rankall), 2026-08-13

xai-value-model/scoring.rs#L52-L56: 'Directly navigating to a post (i.e., coordinating via groupchat) has no ranking impact.' First added in home-mixer/params/param.rs on 14 Aug 2026 (c65aa17), in scoring.rs from 24 Sep 2026. It is written about mass blocking and reporting: 'it must take place on a post served in Home Timeline'. The like-count indexer has no such check in the published code; that is our reading, since the upstream like forwarder is not published. Twitter's 2026 open-sourced code (scoring.rs), 2026-09-24

Twitter's platform manipulation (authenticity) policy permits 'Authentically coordinating with others to express ideas, viewpoints, support, or opposition towards a cause in a non-violative manner', and prohibits 'coordinated inauthentic activity that artificially influences conversations', 'Operating multiple accounts that post substantially similar or identical content ... to the same trending topics or hashtags', creating multiple accounts to 'boost' trending topics or hashtags, and 'Sending bulk, aggressive, high-volume unsolicited replies' (read 7 Nov 2026; the page now redirects to /authenticity, also read in its Wayback copy of 4 Nov 2026). So fans posting a tag together at 9:30 in their own words is within the rules; one person on many accounts is not. What was observed does not show which side of that line every account was on. Twitter Help Center, platform manipulation and authenticity policy

Our researcher's reading of Twitter search's Latest tab for 09:30-09:50 IST on 12 Mar 2026 (read 7 Nov 2026): 232 posts, 101 distinct accounts (2.3 posts each), 98 (42%) only the hashtag, 194 (84%) with a picture, 106 with fan-style names, no default avatars, 8 to 24 posts a minute. Search serves a filtered subset, so these are floors, not totals. Reading (inference): people in a hurry posting content made to be counted, not bots. Not re-checked by the verifier. Twitter search, first 20 minutes of #TNRejectsDMK (our reading), 2026-03-12

The post by @TVKHQITWingOffl, whose bio calls it the official page of TVK's IT wing: created 05:43:25 UTC, text '#TNRejectsDMK' and one image (1,144 reposts, 1,556 likes); four more posts carrying the tag followed. No earlier IT-wing post with the tag was found in archived copies or in Twitter search. #TNRejectsDMK was also the tag of TVK's statewide protest that day. Who briefed whom is not known: the same source could have briefed both, or the wing could have followed the fans. TVK IT wing's official account on Twitter, 2026-03-12

Twitter search of @TVKVijayHQ, @TVKPartyHQ, @TVKHQITWingOffl and @actorvijay for tag, hashtag, trend, their Tamil forms and 'virtual warriors' (read 7 Nov 2026) returned none; absence in search is not proof of absence. The fans' page that set the times says it is 'Not Affiliated With Anyone Officially'. The instructions from Vijay found in the research were about dignity (April 2025) and an offline task (SIR forms, November 2025). Twitter search of TVK's official accounts (our reading) and the fans' page's call

Our researcher's reading in Twitter's default 'Relevant' order, signed in, 7 Nov 2026: first reply shown +0.4 min (recorded as 24 s; said as 'within half a minute' because the second could not be re-read), 22 of 72 inside 5 minutes, 65 inside the hour; 57 of 72 (79%) a picture or video; one default avatar in 215 replies across three posts; a third to a half of the first replies shown carried the paid check. The account is the party president's (@TVKVijayHQ). Reading: an organised, fast, poster-ready fandom, not a bot farm (no repeated text, no egg accounts). Small samples in Twitter's own order; not re-read by the verifier. Vijay's polling-day post on Twitter and its replies (our reading), 2026-04-23

Twitter's page on its paid Premium tiers: Basic 'reply prioritization', Premium 'larger reply prioritization', Premium+ 'largest reply prioritization'; people 'will see a slight preference for replies from verified accounts over other replies' (read 7 Nov 2026). Under TVK's top campaign posts 30-50% of the first replies shown carried the paid check (our sample). Twitter Help Center, About Premium

His Instagram Q&A of 19 May 2024, as quoted by Social Media Today: 'More important than watch time or like and comment counts is send rates'; 'sends per reach correlate more, in my experience, with overall reach than anything else'. His video of 22 Jan 2025, quoted by the same outlet: 'Likes are slightly more important for connected content, and sends are slightly more important for unconnected content.' His words as quoted by a trade outlet; the videos were not re-watched. Meta publishes no weights, so no 'a send = N likes' figure exists. Social Media Today (Adam Mosseri, Instagram Q&A)

@actorvijay reel as Instagram showed it, signed in, 7 Nov 2026: 44,326,742 plays, 3,288,655 likes, 130,159 comments, 328,433 reshares. The star's account, not the party's, carries the reach. Read once by our researcher; Instagram's counts change. Instagram, @actorvijay campaign-song reel (our reading), 2026-04-01

Adam Mosseri, 'Instagram Ranking Explained', 31 May 2023: on Reels, 'Much like Explore, the majority of what you see is from accounts you don't follow'; 'The most important predictions we make are how likely you are to reshare a reel, watch a reel all the way through, like it, and go to the audio page'. Instagram blog, 'Instagram Ranking Explained', 2023-05-31

the-algorithm README (ef4c5eb), line 20: the search index finds 'In-Network Tweets. ~50% of Tweets come from this candidate source'; cr-mixer, the user-tweet graph and follow recommendations fetch out-of-network candidates. A stranger's post also reaches a viewer because people they follow liked it in the last day or two. Twitter's open-sourced code (the-algorithm README), 2023, 2023-03-31

Instagram Reels Chaining AI system card, updated 11 Nov 2025: about 100 reels are scored for each person; predictions include 'How likely you are to share a reel off of Instagram', with the signal 'How many times this post has been shared externally by others', and 'How likely you are to reshare a reel'. No weights are published. That a WhatsApp group sending a reel adds exactly these counts is our inference. Meta Transparency Center, Instagram Reels system card, 2025-11-11

Meta, 6 Aug 2025: the Friends tab shows reels 'your friends have liked, created, reposted or commented on'; a repost 'may be recommended to that person's followers, even if those people don't follow you'. Meta Newsroom, 2025-08-06

Zajonc, 'Attitudinal effects of mere exposure', Journal of Personality and Social Psychology Monograph Supplement 9(2, Pt. 2), 1-27: 'mere repeated exposure of the individual to a stimulus object enhances his attitude toward it'. The faces were yearbook photographs of unknown men, each shown for about 2 seconds a different number of times; 9 of the 12 showed the effect. Read in full at source. It did not study media, timelines, politicians or votes, and it did not show the effect without people remembering (that is a 1980 paper). Journal of Personality and Social Psychology (Zajonc 1968)

Hasher, Goldstein and Toppino, 'Frequency and the conference of referential validity', Journal of Verbal Learning and Verbal Behavior 16(1), 107-112, abstract: 60 plausible statements rated on 'three successive occasions at 2-week intervals'; 'For both true and false statements, there was a significant increase in the validity judgments for the repeated statements and no change in the validity judgments for the non-repeated statements'. It shows 'seems truer', not that people came to believe them. The full text (number of subjects, size of the rise, whether statements were heard or read) was not seen. Journal of Verbal Learning and Verbal Behavior (Hasher, Goldstein and Toppino 1977)

The Hindu, 19 Apr 2025, Vijay in a pre-recorded message: 'Our party IT and social media wing is the best in India. Many have recognised it. You are no longer just fans on social media, you are now TVK's virtual warriors.' He 'urged party cadre to continue maintaining dignity and decency'. In Tamil, in his recorded video message: 'நம்முடைய சோசியல் மீடியா படைதான் இந்தியாவிலேயே பெரிய படைன்னு சொல்றாங்க'. The fans' page used the same words, 'Dear Virtual Warriors', in its 9:30 call. The Hindu, 2025-04-19

New Indian Express, 25 May 2026: an unnamed social media analyst said the campaign deliberately kept content-making away from the IT wing, that the strategists decided content and narrative, and that it was then 'fed to virtual warriors'; a war room of about 40 people with ghostwriters from about June 2025. One outlet, anonymous sources; it says nothing about who set hashtag times. The New Indian Express, 2026-05-25

Times of India, 5 May 2026: 'TVK's official IT wing largely restricted itself to sharing Vijay's speeches'; 'Much of the online mobilisation was driven by supporter accounts and independent digital networks, including IDs funded by Voice of Commons (VoC)', one operative's claim with no evidence shown. An unnamed employee of the firm said they were given the task of 'amplifying' Vijay's complaint of coordinated attacks. 'Funded IDs' is unproven. Times of India, 2026-05-05

Deccan Herald, 26 Aug 2023: at a Vijay Makkal Iyakkam meeting, 1,600 WhatsApp groups with 3 lakh members were to be 'streamlined' and expanded to every constituency, the IT wing to grow to 30,000 people, and members told 'not to indulge in attacking individuals'. The movement's own figure, one outlet. The 34,000 figure (a Tamil news site, Sep 2026, every line 'it is said') names no source. What was sent in the groups, and when, is not documented. Deccan Herald, 2023-08-26

2023 README (b852108): a viewer's negative signal ('show less often', block or mute) -74 (= -148 likes) and report -369 (= -738 likes), against reply 13.5; the reply prediction has no text or tone input. Twitter's 2026 weights (published 13 Aug 2026): not interested -53.07, mute -58.8, block -31.2, report -234, against like 0.708 and reply 5. All are weights on predicted chances, not counts. Twitter's open-sourced ranking code (the-algorithm-ml README), 2023, 2023-04-05

Git history of the 2026 repository: first publication 20 Jan 2026 (aaa167b) with the weights withheld; 13 Aug 2026 (47c1bcd) added the weights, the cold-start lift, the 32-like pool and the label rules; like 0.5 then, 0.708 since 5 Nov 2026; reply 5 (20 between mutual followers). Nothing in it is evidence of what Twitter ran in March and April 2026. Twitter's 2026 open-sourced recommendation code (git history)

grox/flows/reply_spam/task_filter.py#L158-L163 at 3b6919f: the check runs on replies at least two levels deep under a root post whose author has at least 180,000 followers (125,000 until 8 Oct 2026); otherwise the skip reason is 'low_blast_radius'. Real fans replying in their own words with their own pictures trip none of the published copy-paste or bot rules (inference). Twitter's 2026 open-sourced code (reply_spam task filter)

Inference, our arithmetic on the confirmed code: at about 8.3 likes a minute, with re-indexing at most once a minute at the next power of two, the post re-indexes at about 0.1, 1.1 and 2.1 minutes and reaches 32 likes at about 3.8 minutes. It holds only at that pace. The separate cold-start lift counts likes only up to the number of Home timeline views Twitter itself served, so likes from a link do not push that one. Twitter's 2026 open-sourced code (our arithmetic)

Mozilla Foundation report (Odanga Madung and Brian Obilo, Sep 2021): 'added to a Whatsapp group. Here, they received direction about what to post, the hashtags to use ... Synchronizing the tweets was also incredibly important for them. It's what enables them to achieve their goal of trending on Twitter'; 'paid roughly between $10 and $15 to participate in three campaigns per day'; 8 of 11 campaigns reached Trending; 23,606 tweets from 3,742 accounts; the agency selling Twitter ads 'currently offers promoted trends for $3,500 per day'. A foreign case of paid influencers; nothing here is evidence about TVK, and no payment is on record in TVK's case. Mozilla Foundation

Ong and Cabañes, 'Architects of Networked Disinformation' (2018): ad and PR 'chief architects', then anonymous influencers with 50,000 to 2,000,000 followers paid more for more likes, then operators posting scripts 'on a predetermined schedule' to create 'illusions of engagement'. A strategist told them: 'Bots are like the white walkers in Game of Thrones. They're stupid and obvious and easily killed.' Goldsmiths Research Online (Ong and Cabañes 2018)

NATO StratCom COE, 'Social Media Manipulation 2020': 'Of the 337 768 fake engagements purchased, more than 98 per cent remained online and active after four weeks'. Its 2019 test bought 3,530 comments and 25,750 likes for €300. Nothing in the research shows any Tamil Nadu party bought engagement. NATO Strategic Communications Centre of Excellence

Meta's Q2 2023 Adversarial Threat Report, via The Record, 29 Aug 2023: 7,704 Facebook accounts and 954 Pages removed; bursts 'Beijing time, with breaks for lunch and supper'; Page followers mostly from Vietnam, Bangladesh and Brazil. Across the worldwide record, volume without real people flopped and what reached people used real accounts (our synthesis, an inference). The Record (Meta threat report), 2023-08-29

Posted #WeTrustVIJAY 15 minutes after the fans' trends page asked for it in every post; the tag did not reach India's top 50 that day. x.com