# Which Tamil political YouTube channel pulls a newcomer to itself? The method, fixed before any data

Written 17 Nov 2026, about 04:40 IST, before any video of this round is listed, read or scored. Its sha256 and time are
recorded in `../progress.md`. Every change after this point goes into `DEVIATIONS.md` with the time, what changed and
why, before the result it could touch is seen where that is possible. The answer is whatever comes out, whichever channel it
names. Everything done is real and checked, with no guess, assumption or skip, and all the research is published to the
website.

The choices of the method (17 Nov ~04:30 IST): read all the minutes, Opus labelling a random sample and a model on the Mac
learning from it and marking the rest; real AI viewer agents; channels only, chosen by views.

## Why a new round
The round before (v15 to v17) had five faults, said in the published record as they were:
- The "100 AI agents" were 100 rows of random numbers (tools/letdown.py), no model and no browsing.
- 2 to 9 videos a voice, picked as most viewed and recent: 1,212 minutes in all.
- The 23 mixed channels with people heard on other channels.
- Marks were places within the 23 (0 = the lowest, 100 = the highest), not amounts.
- The agents' bar (30 to 50) was assumed, and was changed after a first result was seen. The winner (Savukku) came
  first in 53% of redraws.

## 1. The channels (fixed by rule)
- **Source:** this month's political videos in YouTube's own search sorted by views (`../research/ml/out/tier.json`,
  1,133 videos, made 16 Nov before any scoring).
- **In by views:** every channel of kind commentator, interview_platform or digital_news with 1,000,000 or more
  political views there.
  - Out: TV news, fan or clip channels, comedy, parties and politicians, "other" and entertainment.
  - Out: collaboration labels that are not a channel of their own (YouTube's "X and Y" names), which are counted
    with their channel.
- **In by name:** the channels written into this plan by name (Saattai, Maridhas Answers, U2 Brutus, Idam Valam, AranSei,
  Tribes, Galatta Voice, Aadhan News, Red Pix 24x7), whatever their views.
- **The list (20):**
  - By views: Savukku Media Network, Arasiyal Sadhurangam, Saattai, Chanakyaa, Minnambalam, Neerthirai, Aadhan News,
    Behindwoods O2, Idam Valam, Dream Big with SK, RETROLUXE, Maridhas Answers, info politics Venkee Rathinam,
    Vikatan TV, IBC Tamil.
  - By name: U2 Brutus, AranSei, Tribes, Galatta Voice, Red Pix 24x7.
- A channel's kind is the label given on 16 Nov (`tools/instructions-channels.md`) and is published with it.
- People heard on other channels are not units. The satire channel the show never names is not on the list.

## 2. The videos (all of them, in the window)
- **The window:** every upload from 17 Sep 2026 00:00 IST to 16 Nov 2026 23:59 IST, listed from each channel's
  Videos, Shorts and Live tabs (yt-dlp).
  - Each upload's date is read from its own page (not the listing's approximate date).
  - A listing cut short by a cap is continued until the window is covered.
- **Political:**
  - A video is political when it is about Tamil Nadu or Indian politics: parties, leaders, government, elections,
    policy, or politics-related public affairs.
  - Each title and description are read by Opus readers (two readers on every title, a third on a disagreement).
  - The decision and the reason are published for every video.
- **The main result** is on long videos and Shorts.
  - Live streams are listed, counted and scored the same way where they have captions, and reported as a second table.
  - Live streams are never mixed into the main result.
- **Captions:**
  - The text is YouTube's own captions, Tamil first: the creator's if there are any, else the automatic ones.
  - A video with no captions is listed with its length, never dropped silently. The share of each channel's minutes
    without text is published beside its marks.

## 3. The minutes
- Each political video's captions are cut into 60-second windows from 0 (the last window kept when 20 s or longer).
  A Short is one window.
- Every window gets an id (video id + start) and its text.

## 4. The marks: what a minute does, in plain amounts
Each label is true or false for a minute, read from its text. A channel's mark is the share of its political minutes
where the label is true, in percent ("out of 100 minutes, so many"), with a 95% interval from a bootstrap over its
videos. No rescaling to the other channels.

The six labels, with the reading instructions written in `instructions-labels.md` before the sample is drawn:
1. **Detail** (விவரம்): the minute gives at least one specific: a number, a date, a document, or a named source.
2. **Questions the state government** (மாநில அரசை கேள்வி கேக்குது): the minute criticises or questions the Tamil
   Nadu government (the T.V.K government since May 2026), its ministers or its decisions.
3. **Questions the Union government** (ஒன்றிய அரசை கேள்வி கேக்குது): the same for the Union government (BJP-led).
4. **Shows proof** (ஆதாரம் காட்டுது): the minute backs a claim with something checkable, shown or cited (a document,
   a record, a clip, a named report), not only an assertion.
5. **Makes you laugh** (சிரிக்க வைக்குது): the minute is meant to be funny (jokes, mockery, satire).
6. **Stirs feeling** (உணர்ச்சி): the minute is charged with anger, fear, pride or ridicule.

Also recorded for every minute, not a mark: which party or leader it attacks or favours (as before:
`../research/tools/instructions-sides.md`).

## 5. Who reads the minutes
- **The Opus sample:**
  - 3,000 windows drawn at random with seed 20261117, stratified by channel (150 a channel, all of a channel's windows
    when it has fewer).
  - Within a channel every political window has the same chance.
  - Read by Opus subagents on `instructions-labels.md`.
  - A second Opus reader reads 300 of them (drawn with seed 20261118) blind to the first. Agreement (Cohen's kappa)
    is published for every label.
- **The model on the Mac:**
  - Every window's text is embedded (multilingual-e5-large, as in round 1).
  - A logistic regression is trained for each label on 80% of the sample. 20% (seed 20261119) is held out, never
    trained on or tuned on.
  - The 1,212 windows read in rounds 1 to 5 may be added to the training data only, never to the held-out set and
    never to a channel's mark.
  - On the held-out set, each label's precision, recall and agreement with Opus (kappa) are published.
- **The bar for using the model:**
  - Where a label's held-out kappa with Opus is 0.60 or more (about where the two Opus readers agree), the model
    marks every window. The channel's share is then corrected for the model's measured errors on the held-out set
    (adjusted classify-and-count) and published with both numbers.
  - Where a label falls under 0.60, the model is not used for it. That label's mark is the Opus sample's share, with
    its interval, and that is said.

## 6. The viewer agents (real AI agents)
- **Who:** 100 Opus agents, each one newcomer to Tamil Nadu politics on YouTube with no party leaning (a viewer
  who is interested in Tamil Nadu politics without a prior bias).
  - The profiles differ in age, gender, place (city, town, village) and what they want from a political video.
  - They are drawn in proportion to published figures (Tamil Nadu's adult population by age and gender, Census and
    the 2026 roll; urban and rural, Census 2011; issues voters named, the 2026 post-poll surveys). Each figure is cited
    in `profiles.md`.
  - The 100 profiles are written and published before any agent reads.
- **What they read:**
  - The same 8 random minutes from each channel (seed 20261120, from its political long videos and Shorts in the
    window), as transcript text.
  - Channel names are replaced by letters, shuffled for each agent; the channels come in a random order for each agent.
  - Nobody browses: YouTube served this machine a feed for its own country, not India (16 Nov), so a browsing agent
    here would not see what a viewer in Tamil Nadu sees.
- **What they answer:**
  - Which channels they would follow, up to three, ranked, with one line why for each.
  - For every channel: would it pull them toward a party, and which.
  - Whether they recognised any channel or speaker, and which.
- **Records:** the answers are stored whole and published.
- **The result:**
  - The channel picked by the most agents among their three (ties said as ties), with first choices beside it, and
    the interval from a bootstrap over the agents.
  - The marks of part 4 are then read against the agents' choices to say what the chosen channels do differently.
    That is a description, not a cause.

## 7. What the video may say
- Only what parts 1 to 6 show.
- Marks are said as amounts ("நூறு நிமிஷத்துல இத்தனை").
- The agents are said as what they are: AI agents that read the same random minutes, not people and not a poll.
- Every limit that changes a reading is said in the video or on its page.
- No comment's words and no commenter's name anywhere.

## 8. What is published (politic.in, the issue's research)
- This plan.
- `DEVIATIONS.md`.
- The channel list and its rule.
- Every listed upload with its date, length, political decision and reason, and caption status.
- Every window's labels: Opus's and the model's.
- The agreement and the held-out scores.
- The code.
- The 100 profiles and the 100 answers.
- The results with intervals.
- What the round before got wrong.

Before anything goes live, the page is reviewed.
