# Which Tamil political YouTube channel pulls a newcomer to itself? The record of this round

> Where a figure in the body below was later corrected, the correction is in the section "Corrections after a later review" (section 8) and in `DEVIATIONS.md`; the corrected figure is the one that stands.

This is a plain account of how this round was measured: every count, every change to the method and every limit. The
method was written and fixed before any data of the round was read (`PLAN.md`, sha256
693d4ad196876039771ad83af44ad67f80e6e4e29ef6313432471631349e8720; the published copy has a few lines reworded, listed in `README.md`). Every change after that is in `DEVIATIONS.md`,
with its time and its reason. No channel was meant to win: the answer is whatever the data gives.

## What the round before got wrong
The round before this one had five faults. They are said here as they were:
- **The "100 AI agents" were not agents.** They were 100 rows of random numbers: each row was a random mix of what an
  agent cares about, plus a random bar. No model read anything and nothing was browsed.
- **Too little was read.** It read 2 to 9 videos a voice, picked as the most viewed and the most recent: 1,212
  minutes in all.
- **The units were mixed.** Its 23 "voices" mixed channels with people who are heard on other channels.
- **The marks were ranks, not amounts.** A mark was a channel's place among the 23 (0 = the lowest, 100 = the
  highest), not how much of something it does.
- **The bar was assumed, and moved after a result.** The agents' bar (30 to 50) was assumed, not measured. It was
  changed after a first result was seen. The channel that came first (Savukku Media Network) came first in only 53% of
  redraws.

This round fixes each of them:
- The viewer agents are real AI agents that read text.
- The minutes are read whole, from far more videos.
- The units are channels only.
- The marks are amounts ("out of 100 minutes, so many").
- No bar is assumed.

## The question and the window
The question: if someone new to Tamil Nadu politics, with no leaning, starts watching Tamil political YouTube, which
channel would pull them to itself?

The window: uploads from 17 Sep 2026 00:00 IST to 16 Nov 2026 23:59 IST. An upload's date is the one on its own page
(the start of the stream for a stream).

## 1. The channels (21)
The channels were chosen by a rule fixed before any of their videos was read:
- **In by views:** every channel whose kind is commentator, interview platform or digital news, with 1,000,000 or
  more political views in YouTube's own search sorted by views for this month (a sweep of 1,133 political videos and
  247 channels, made on 16 Nov before any scoring). A collaboration ("X and Y") counts with the channel that uploaded
  it, read from the video's own page. That rule added Reflect News Tamil, which the plan's written list of 20 had
  missed (DEVIATION 2).
- **Out:** TV news channels, fan and clip channels, comedy, parties and politicians, entertainment. People who are
  heard on other channels are not units of their own. A satire channel that this show does not name is not on the
  list.
- **In by name:** the nine channels written into the plan by name (Saattai, Maridhas Answers, U2 Brutus, Idam Valam, AranSei,
  Tribes, Galatta Voice, Aadhan News, Red Pix 24x7), whatever their views. Four of them are in by views as well, so
  five are in by name only.

| Channel | Kind | In by | Political views in the sweep | Channel id |
|---|---|---|---:|---|
| Savukku Media Network | commentator | views | 11,786,468 | UC3MiyqRMvE5GZm8mqmbOQBg |
| Arasiyal Sadhurangam | commentator | views | 8,305,517 | UCsWipAXDjOrnfiCGoA1zSAA |
| Chanakyaa | commentator | views | 5,296,890 | UCoOu4D7foJWfKvcDLxqrF2Q |
| Saattai | commentator | views | 4,005,776 | UCyQEgd5bpLWVkFs3-fdYMQQ |
| Minnambalam | digital news | views | 3,492,605 | UCgFSoS8vu0ONak4z5OBORHw |
| Neerthirai | digital news | views | 3,479,600 | UCiQqeecBr38vyQCTq-ef9MQ |
| Aadhan News | interview platform | views | 3,279,149 | UCkscWUgDcYN1tSliZkLxdBg |
| Behindwoods O2 | interview platform | views | 3,178,951 | UC38z3fT9RO4yugLJoCZLygw |
| Idam Valam | commentator | views | 2,858,083 | UCvU9vF3a3QDMljWim5ztBPA |
| Dream Big with SK | commentator | views | 2,346,392 | UCTLbDyGIIrcGNzc-oG8dkSg |
| RETROLUXE | interview platform | views | 2,073,093 | UCrdaqAg302_SuXhumzxDEXg |
| Maridhas Answers | commentator | views | 1,956,133 | UCN0lfcRcYiMzLzGn1SnSsXw |
| info politics Venkee Rathinam | commentator | views | 1,337,569 | UCKjGXjQ-M5DdjCRWDqdtjjA |
| Vikatan TV | digital news | views | 1,106,037 | UCEhsOgK2u8GDDoynMCj78Ng |
| IBC Tamil | digital news | views | 1,088,095 | UCZinho9wu3iAeRpUXvAMzBQ |
| Reflect News Tamil | interview platform | views | 1,040,909 | UC9gcI3j2_p1GS-RlHtl2_wg |
| U2 Brutus | commentator | name | 0 | UCBN-Kg9WxabjMgw4yHLLV7g |
| AranSei | commentator | name | 0 | UCQvcXzhYS0X-fH6PjVGEBDw |
| Tribes | commentator | name | 273,226 | UCe4Xx2aBYBTncbWnmw5mg6A |
| Galatta Voice | interview platform | name | 153,172 | UCUWdltdSR6mONwLa5ODDVjw |
| Red Pix 24x7 | interview platform | name | 498,283 | UCe1Nb-N6XgLFHfGxZjIkF8A |

The sweep's views are those of the political videos it found, not each channel's whole count.

## 2. Which videos were read: chosen by views, not all
The plan was to read every upload of the window. The round had to end in about six hours, so the videos were chosen
by views instead (DEVIATION 1): for each channel, its most-viewed political uploads in the window, up to 10 long
videos and 5 Shorts. Live streams are left out of this round.

How a channel's uploads were gone through:
1. **Listed.** Each channel's Videos and Shorts tabs were listed with their view counts. A listing's views are
   rounded (for example "91K") and its dates approximate. So no upload was kept or dropped on a listing's date alone.
2. **Ranked by views.** Within each tab, uploads were taken in order of listed views, highest first.
3. **Title screen.** Two AI readers (Opus) read the title of every upload high in a tab's ranking, each on its own:
   political, not political, or cannot tell. An upload that BOTH readers found clearly not political was not opened.
4. **Opened.** Every other upload had its own page read once: its exact date, its length, whether it was a live
   stream, its description and its captions, all in the same visit.
5. **Read as political or not.** Two fresh AI readers (Opus) read its title and full description, each on its own; a
   third read it where they differed. Only this reading keeps a video.
6. **Kept.** Going down the ranking, an upload was kept when its own page dates it inside the window, it is not a
   live stream, and the readers found it political, until the channel had 10 long videos and 5 Shorts, or ran out.

Every upload gone through is published with what happened to it and why.

**How far the reading got.** YouTube refused this round's page reads once (HTTP 429 at 04:41 IST, after about 218
pages in 12 minutes). After that, pages were read one at a time, about one every 10 to 11 seconds (DEVIATIONS 3 and
5), and a last time was set for page visits: 06:45 IST (DEVIATION 7). The visits ended at 06:42 IST, before that time,
because nothing was left to open, so the time limit cut no tab.
- 38 of the 42 tabs (21 channels, long videos and Shorts) reached their 10 or 5.
- The other four are Shorts tabs (Idam Valam, Maridhas Answers, Red Pix 24x7, info politics Venkee Rathinam): the
  newest Short on each is dated before 17 Sep 2026 on its own page, so they have no Shorts in the window.
- In every tab, every upload ranked above the last kept one has a final state. The table in section 6 says, for each
  tab, how far down its view ranking that holds.
- One fault in the planning of the page visits was found and mended before the kept videos were fixed: the planner
  took a tab to be full as soon as 10 uploads were marked kept, even with uploads above the tenth not yet opened. In
  Idam Valam's long tab, ranks 10 to 12 had not been opened; they were opened and read, and the tab's ten kept videos
  are its ranks 1 to 10 (the note under DEVIATION 7).

## 3. The text and the minutes
- **The text** is YouTube's own captions: the creator's Tamil captions where there are any, else YouTube's automatic
  Tamil captions. Some automatic tracks are labelled with another language on YouTube (for example Korean) while the
  text in them is Tamil; they are used as Tamil text, and their label is recorded.
- A kept video with no text is listed with its length and never dropped. The share of each channel's minutes with no
  text is published beside its marks.
- **The minutes:** each kept video's text is cut into 60-second windows from 0. The last window is kept when it is 20
  seconds or longer (by the video's length). A Short is one window. Every window has an id (video id + start second).

## 4. What a minute does: the six marks
Every minute with text was read by an AI reader (Opus) that saw only the minute's text: no channel, no title, no
picture, and minutes of many channels mixed in one packet. For each minute it said true or false for:
1. **Detail** (விவரம்): a number, a date, a document or a named source.
2. **Questions the state government** (மாநில அரசை கேள்வி கேக்குது).
3. **Questions the Union government** (ஒன்றிய அரசை கேள்வி கேக்குது).
4. **Shows proof** (ஆதாரம் காட்டுது): something checkable, shown or cited, not only said.
5. **Makes you laugh** (சிரிக்க வைக்குது).
6. **Stirs feeling** (உணர்ச்சி): anger, fear, pride or ridicule.
It also said whether the text was readable at all, and which party the minute attacks or favours (not a mark).

A channel's mark is the share of its readable minutes (minutes with text that the reader found readable; DEVIATION 6)
where the label is true: "out of 100 minutes, so many". Nothing is
rescaled against the other channels. Each mark has a 95% interval from a bootstrap over the channel's videos (2,000
draws), because minutes of one video are alike. A second column weights each video by its views.

**How far two readers agree.** 300 minutes, drawn at random, were read again by a second AI reader (Opus) who did not
see the first reading. Cohen's kappa for each label is published.

**What was not done.** The plan was to have the AI readers read a sample and to train a model to mark the rest. With
the videos chosen by views, every kept minute could be read by the AI readers themselves, so no model was trained,
and no model's scores apply (DEVIATION 1).

## 5. The viewer agents
- **Who:** 100 AI agents (Opus). Each plays one newcomer to Tamil Nadu politics on YouTube, with no party leaning.
  - The 100 differ in age, gender, place (city, town or village) and what they want a political video to explain.
  - Each trait is drawn in proportion to a published figure: Census 2011 for age, gender and place; a 2021 post-poll
    survey for what voters named as their issue (DEVIATION 4 says why not the 2026 roll and a 2026 survey).
  - The 100 profiles were fixed in `profiles.md` before any agent read anything (they are published with this record).
- **What they read:** the same 8 random minutes from each channel, as text.
  - The minutes are drawn with a fixed seed (20261120) from the minutes of the channel's kept long videos and Shorts
    that the first reader found readable (DEVIATION 8). Two drawn minutes that speak of a creator's arrest or legal
    case were replaced by other minutes of the same channel (DEVIATION 9).
  - Each channel is called by a letter in place of its name; the minutes are left as spoken, so some still say a name
    (see "Names said aloud" in section 8). Each agent has its own letters and its own order of channels.
  - No agent browses YouTube: a browser here would be shown another country's feed, not India's.
- **What they answer:**
  - which channels they would follow (up to three, ranked, one line why for each);
  - for every channel, whether it would pull them toward a party, and which;
  - whether they recognised any channel or speaker.
- The agents are AI agents that read the same random minutes. They are not people, and their answers are not a poll.

## 6. What was read, channel by channel
Every count below is made from the published files. "Read down to rank" is how far down the tab's view ranking every
upload has a final state (kept, or out, with the reason). "Share of window views" is the kept videos' listed views as
a share of the listed views of all the channel's uploads in the window, political or not; it is a range where a
listing's dates did not fix the window's edges exactly.

| Channel | Long kept | Shorts kept | Read down to rank (long / Shorts) | Minutes | With text | Readable | Kept views (listed) | Share of window views |
|---|---:|---:|---|---:|---:|---:|---:|---|
| Savukku Media Network | 10 | 5 | 10 / 29 | 602 | 602 | 602 | 17,653,000 | 27% to 30% |
| Arasiyal Sadhurangam | 10 | 5 | 10 / 67 | 189 | 186 | 184 | 5,056,000 | 21% to 22% |
| Chanakyaa | 10 | 5 | 11 / 6 | 208 | 207 | 206 | 9,223,000 | 14% to 16% |
| Saattai | 10 | 5 | 14 / 32 | 289 | 289 | 266 | 11,402,000 | 14% to 15% |
| Minnambalam | 10 | 5 | 13 / 5 | 143 | 143 | 142 | 11,728,000 | 28% to 34% |
| Neerthirai | 10 | 5 | 23 / 29 | 299 | 297 | 290 | 7,255,000 | 21% to 29% |
| Aadhan News | 10 | 5 | 15 / 8 | 439 | 439 | 438 | 8,711,000 | 21% to 26% |
| Behindwoods O2 | 10 | 5 | 14 / 9 | 259 | 259 | 254 | 9,171,000 | 24% to 26% |
| Idam Valam | 10 | 0 | 10 / 578 (none in window) | 226 | 226 | 191 | 5,421,000 | 50% |
| Dream Big with SK | 10 | 5 | 10 / 7 | 180 | 180 | 180 | 6,187,000 | 30% to 41% |
| RETROLUXE | 10 | 5 | 19 / 7 | 325 | 325 | 325 | 4,576,000 | 16% to 20% |
| Maridhas Answers | 10 | 0 | 15 / 75 (none in window) | 134 | 134 | 134 | 4,466,000 | 32% |
| info politics Venkee Rathinam | 10 | 0 | 11 / 53 (none in window) | 164 | 164 | 151 | 2,716,000 | 28% |
| Vikatan TV | 10 | 5 | 27 / 18 | 267 | 266 | 264 | 2,996,000 | 11% to 12% |
| IBC Tamil | 10 | 5 | 43 / 12 | 372 | 371 | 369 | 2,742,000 | 13% |
| Reflect News Tamil | 10 | 5 | 10 / 23 | 60 | 60 | 58 | 4,763,000 | 35% to 38% |
| U2 Brutus | 10 | 5 | 11 / 84 | 304 | 304 | 304 | 4,421,000 | 22% to 26% |
| AranSei | 10 | 5 | 14 / 15 | 341 | 341 | 340 | 1,768,000 | 26% to 27% |
| Tribes | 10 | 5 | 10 / 84 | 370 | 370 | 370 | 4,120,000 | 28% to 29% |
| Galatta Voice | 10 | 5 | 13 / 18 | 236 | 233 | 231 | 7,029,000 | 25% |
| Red Pix 24x7 | 10 | 0 | 11 / 441 (none in window) | 356 | 356 | 321 | 1,897,000 | 44% |

*Added 17 Nov 2026, 16:27 IST:* the last column rests on the first Shorts listings, which for six channels did not
reach the window's start. With the deeper listings (DEVIATION 16) it is: Galatta Voice 7.4%, IBC Tamil 6.1% to 6.2%,
Behindwoods O2 19.7% to 19.9%, Chanakyaa 13.7% to 14.2%, Minnambalam 29.3% to 30.2%, Vikatan TV 10.9% to 11.9%; the
other fifteen are unchanged. See "Corrections after a later review" in section 8.

- **All together:** 1,914 uploads gone through, 417 pages opened, 330 uploads read on their title and description by
  two readers (4 of them by a third, where the two differed), 295 kept: 210 long videos and 85 Shorts.
- **The text:** none of the 295 kept videos had Tamil captions made by the creator. 228 have YouTube's automatic Tamil
  track; 64 have an automatic track that YouTube labels with another language (Korean 47, English 11, English and
  Korean 4, US English and Korean 1, Hindi 1); 3 Shorts have no captions at all (Chanakyaa, 15 s; Galatta Voice, 27 s;
  IBC Tamil, 22 s).
- **The minutes:** 5,763 in all, 5,752 with text, 11 without (the 3 Shorts above and 8 minutes of long videos where
  the captions have a gap: Neerthirai 2, Arasiyal Sadhurangam 3, Galatta Voice 2, Vikatan TV 1). The first reader
  found 5,620 of the 5,752 readable.
- **The kept videos hold between 22% and 25% of the listed views of the 21 channels' uploads in the window**
  (133,301,000 of between 543,138,979 and 613,762,070).
  *Corrected 17 Nov 2026, 16:27 IST:* with the deeper Shorts listings (DEVIATION 16), between 18.8% and 20.2%
  (133,301,000 of between 658,386,233 and 710,490,828).

## 7. The marks
Each cell is "out of 100 readable minutes of the channel, so many", with its 95% interval (bootstrap over the
channel's videos, 2,000 draws, seed 20261117). The last two columns are the shares of the channel's minutes without
text and of its minutes with text that the reader found unreadable; neither is in the marks. A channel with few
videos or few minutes (Reflect News Tamil: 58 readable minutes) has wide intervals.

| Channel | Readable minutes | Detail | Questions state govt | Questions Union govt | Shows proof | Makes you laugh | Stirs feeling | No text | Unreadable |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| Savukku Media Network | 602 | 63.8 (57.1-70.9) | 43.4 (32.4-53.2) | 0.0 (0.0-0.0) | 23.3 (19.4-27.6) | 42.7 (33.5-52.0) | 57.5 (50.5-63.4) | 0.0% | 0.0% |
| Arasiyal Sadhurangam | 184 | 15.2 (9.8-22.0) | 3.3 (0.7-5.9) | 0.0 (0.0-0.0) | 6.0 (3.3-9.1) | 37.5 (28.4-48.2) | 59.8 (48.6-69.4) | 1.6% | 1.1% |
| Chanakyaa | 206 | 61.7 (45.8-78.9) | 12.6 (4.9-23.8) | 0.5 (0.0-1.6) | 21.4 (1.7-47.5) | 4.4 (1.9-6.8) | 8.3 (3.4-14.8) | 0.5% | 0.5% |
| Saattai | 266 | 53.8 (43.1-65.3) | 40.2 (29.3-48.7) | 0.4 (0.0-1.3) | 6.4 (1.0-15.3) | 50.8 (35.6-65.6) | 74.4 (64.5-84.2) | 0.0% | 8.0% |
| Minnambalam | 142 | 35.9 (27.7-45.0) | 26.8 (10.5-46.7) | 9.9 (1.4-22.1) | 2.8 (0.9-4.8) | 23.2 (9.0-38.7) | 45.8 (28.8-60.7) | 0.0% | 0.7% |
| Neerthirai | 290 | 55.2 (43.3-62.9) | 46.6 (23.5-75.2) | 17.2 (0.8-32.5) | 5.2 (0.0-10.0) | 34.8 (22.5-51.7) | 80.3 (71.6-92.8) | 0.7% | 2.4% |
| Aadhan News | 438 | 39.5 (28.8-50.5) | 19.6 (10.7-29.0) | 1.8 (0.4-4.1) | 2.1 (0.9-3.3) | 39.5 (26.5-52.6) | 69.9 (52.9-82.3) | 0.0% | 0.2% |
| Behindwoods O2 | 254 | 49.6 (40.1-58.1) | 13.8 (4.7-26.4) | 0.8 (0.0-2.5) | 3.1 (0.9-5.7) | 10.6 (2.5-19.7) | 16.9 (8.4-26.4) | 0.0% | 1.9% |
| Idam Valam | 191 | 37.7 (27.3-46.0) | 7.3 (2.0-13.7) | 0.5 (0.0-1.5) | 10.5 (4.5-17.5) | 37.2 (20.2-49.8) | 68.6 (48.8-83.1) | 0.0% | 15.5% |
| Dream Big with SK | 180 | 86.7 (80.3-91.8) | 11.1 (1.5-27.0) | 1.7 (0.0-5.5) | 39.4 (28.0-52.8) | 9.4 (3.7-15.4) | 33.9 (20.3-48.4) | 0.0% | 0.0% |
| RETROLUXE | 325 | 58.2 (47.8-69.2) | 11.4 (2.9-22.6) | 3.1 (0.0-7.8) | 9.2 (3.3-17.2) | 21.8 (9.3-41.0) | 46.2 (32.7-58.6) | 0.0% | 0.0% |
| Maridhas Answers | 134 | 65.7 (57.6-73.8) | 32.8 (18.8-49.6) | 0.0 (0.0-0.0) | 8.2 (1.5-16.3) | 8.2 (2.8-14.3) | 35.1 (18.8-53.8) | 0.0% | 0.0% |
| info politics Venkee Rathinam | 151 | 41.1 (31.8-50.3) | 13.2 (5.1-23.7) | 0.0 (0.0-0.0) | 4.6 (1.9-7.7) | 43.0 (25.5-59.1) | 73.5 (56.7-86.0) | 0.0% | 7.9% |
| Vikatan TV | 264 | 49.6 (37.3-64.7) | 20.5 (8.1-32.9) | 3.0 (0.5-6.0) | 4.2 (2.1-6.8) | 18.6 (9.6-30.5) | 28.8 (12.1-47.2) | 0.4% | 0.8% |
| IBC Tamil | 369 | 48.8 (42.6-57.3) | 21.4 (6.3-34.0) | 3.5 (0.0-6.2) | 3.0 (0.0-4.7) | 15.2 (2.0-26.9) | 30.9 (9.3-48.6) | 0.3% | 0.5% |
| Reflect News Tamil | 58 | 25.9 (8.0-34.0) | 5.2 (0.0-27.6) | 0.0 (0.0-0.0) | 3.4 (0.0-9.7) | 32.8 (3.6-43.1) | 63.8 (31.0-75.4) | 0.0% | 3.3% |
| U2 Brutus | 304 | 31.9 (20.3-45.1) | 24.0 (11.4-38.0) | 2.0 (0.4-3.9) | 14.5 (7.6-22.1) | 44.4 (31.4-61.3) | 71.7 (63.1-82.9) | 0.0% | 0.0% |
| AranSei | 340 | 52.9 (40.7-67.4) | 30.9 (16.6-47.4) | 17.4 (2.4-30.4) | 11.5 (5.6-19.3) | 15.0 (9.1-20.3) | 31.8 (16.4-45.2) | 0.0% | 0.3% |
| Tribes | 370 | 56.5 (47.0-64.0) | 29.5 (17.2-44.7) | 32.7 (10.4-51.6) | 19.5 (13.1-25.0) | 20.3 (13.0-30.4) | 64.9 (53.8-78.9) | 0.0% | 0.0% |
| Galatta Voice | 231 | 40.7 (31.5-50.5) | 19.5 (10.2-28.1) | 6.5 (0.0-16.4) | 2.2 (0.5-4.1) | 19.5 (12.0-27.4) | 40.3 (26.1-53.0) | 1.3% | 0.9% |
| Red Pix 24x7 | 321 | 52.3 (33.7-71.3) | 10.3 (4.0-16.7) | 0.3 (0.0-1.0) | 14.3 (4.2-26.4) | 9.7 (4.3-15.8) | 17.8 (8.2-28.3) | 0.0% | 9.8% |

**Weighted by views.** The same marks with each video weighted by its views (the count on its own page when read), so
a channel's most-watched videos count for more:

| Channel | Detail | Questions state govt | Questions Union govt | Shows proof | Makes you laugh | Stirs feeling |
|---|---:|---:|---:|---:|---:|---:|
| Savukku Media Network | 68.8 | 36.7 | 0.0 | 22.4 | 40.4 | 63.0 |
| Arasiyal Sadhurangam | 20.1 | 2.9 | 0.0 | 6.3 | 37.5 | 58.1 |
| Chanakyaa | 66.7 | 7.9 | 0.3 | 16.1 | 2.6 | 16.2 |
| Saattai | 59.2 | 39.4 | 0.4 | 4.6 | 65.7 | 81.6 |
| Minnambalam | 66.8 | 35.1 | 18.7 | 1.7 | 14.0 | 35.8 |
| Neerthirai | 19.0 | 72.6 | 3.6 | 1.0 | 27.7 | 93.6 |
| Aadhan News | 36.2 | 21.7 | 0.8 | 1.5 | 40.1 | 83.5 |
| Behindwoods O2 | 46.7 | 16.8 | 0.5 | 2.1 | 7.3 | 11.0 |
| Idam Valam | 36.5 | 7.3 | 0.4 | 9.9 | 35.1 | 70.3 |
| Dream Big with SK | 95.7 | 3.3 | 0.6 | 55.8 | 18.4 | 68.0 |
| RETROLUXE | 57.1 | 8.2 | 2.9 | 8.8 | 31.2 | 57.1 |
| Maridhas Answers | 65.7 | 28.4 | 0.0 | 9.5 | 7.8 | 33.4 |
| info politics Venkee Rathinam | 41.4 | 10.8 | 0.0 | 3.8 | 33.0 | 61.4 |
| Vikatan TV | 54.1 | 16.7 | 2.1 | 14.5 | 17.8 | 25.3 |
| IBC Tamil | 49.2 | 21.6 | 1.7 | 1.1 | 6.9 | 30.9 |
| Reflect News Tamil | 11.2 | 5.6 | 0.0 | 2.7 | 10.4 | 59.6 |
| U2 Brutus | 51.4 | 26.8 | 0.7 | 18.4 | 38.3 | 70.7 |
| AranSei | 53.8 | 35.6 | 12.9 | 15.2 | 18.9 | 34.0 |
| Tribes | 49.1 | 49.4 | 15.6 | 15.6 | 41.5 | 82.6 |
| Galatta Voice | 37.0 | 16.5 | 6.2 | 1.6 | 18.7 | 33.8 |
| Red Pix 24x7 | 52.0 | 9.2 | 0.2 | 16.0 | 9.2 | 18.8 |

**How far two readers agree.** 300 minutes with text, drawn at random (seed 20261118) from all 21 channels, were read
again by a second AI reader (Opus) who did not see the first reading. Both readers are the same model with the same
instructions, so this measures how steady the reading is, not whether a person would agree.

| Label | Cohen's kappa | Raw agreement | True for the first reader | True for the second reader |
|---|---:|---:|---:|---:|
| Detail | 0.887 | 94.3% | 51.0% | 52.0% |
| Questions the state government | 0.816 | 93.3% | 23.0% | 24.3% |
| Questions the Union government | 0.838 | 97.7% | 7.3% | 8.3% |
| Shows proof | 0.792 | 95.7% | 12.7% | 11.0% |
| Makes you laugh | 0.881 | 95.3% | 27.7% | 25.7% |
| Stirs feeling | 0.851 | 92.7% | 44.7% | 44.0% |

Both readers found the same 296 of the 300 minutes readable (kappa 1.0 on "readable").

The marks say what a channel's minutes do. They do not say which channel pulls a newcomer: that is the viewer agents'
part (section 5), read against these marks afterwards.

## 8. What the viewer agents chose, and the check that changed the answer
**Draw 1 (the plan).** 100 agents answered; 99 are counted (P069 wrote "toward TVK" where the form asks "TVK", four
times, and the rule written before any answer leaves an answer out of form; DEVIATION 12). Every one of the 99 named
up to three channels; none chose none (96 named three, 3 named two). Vikatan TV was in the three of all 99 and first for
93. With P069 counted, Vikatan TV is in 100 of 100.

**The check (DEVIATION 13, written after draw 1's result and before the check ran).** Every agent in draw 1 read the
same 8 minutes of each channel, so draw 1's intervals are over the agents only, not over which minutes were drawn, and
Vikatan TV's 8 drawn minutes had none that stirred feeling while 28.8% of all its readable minutes do. So 20 of the
100 profiles (drawn with seed 20261126) read a fresh 8 minutes of each channel (seed 20261125, none from draw 1), with
the same instructions, letters and order. All 20 answered in form.

**The answer changed.** In draw 2, RETROLUXE was in 12 of the 20 agents' three and first in 69% of the redraws;
Vikatan TV was in none. By the rule written before the check, **the leader is not settled: which channel the agents
pick depends on which minutes they read.**

| Channel | Draw 1: in the three (of 99) | 95% interval | first choice | Draw 2: in the three (of 20) | 95% interval | first choice | Feeling in the 8 read, draw 1 / draw 2 |
|---|---:|---|---:|---:|---|---:|---|
| Vikatan TV | 99 | 99-99 | 93 | 0 | 0-0 | 0 | 0 / 5 |
| AranSei | 77 | 69-85 | 3 | 5 | 2-9 | 3 | 1 / 2 |
| Chanakyaa | 52 | 42-62 | 0 | 3 | 0-6 | 0 | 0 / 0 |
| Dream Big with SK | 13 | 7-20 | 1 | 8 | 4-12 | 4 | 3 / 2 |
| Savukku Media Network | 13 | 7-20 | 0 | 3 | 0-6 | 3 | 3 / 1 |
| Behindwoods O2 | 12 | 6-19 | 0 | 9 | 5-13 | 3 | 2 / 1 |
| Neerthirai | 11 | 5-17 | 1 | 0 | 0-0 | 0 | 5 / 7 |
| Red Pix 24x7 | 10 | 5-16 | 1 | 9 | 5-13 | 1 | 2 / 2 |
| Minnambalam | 6 | 2-11 | 0 | 0 | 0-0 | 0 | 4 / 5 |
| U2 Brutus | 1 | 0-3 | 0 | 1 | 0-3 | 0 | 7 / 5 |
| RETROLUXE | 0 | 0-0 | 0 | 12 | 8-16 | 5 | 4 / 0 |
| Galatta Voice | 0 | 0-0 | 0 | 4 | 1-8 | 0 | 3 / 5 |
| IBC Tamil | 0 | 0-0 | 0 | 4 | 1-8 | 0 | 4 / 1 |
| Saattai | 0 | 0-0 | 0 | 1 | 0-3 | 0 | 8 / 5 |
| Tribes | 0 | 0-0 | 0 | 1 | 0-3 | 1 | 5 / 4 |
| Aadhan News | 0 | 0-0 | 0 | 0 | 0-0 | 0 | 6 / 7 |
| Arasiyal Sadhurangam | 0 | 0-0 | 0 | 0 | 0-0 | 0 | 4 / 5 |
| Idam Valam | 0 | 0-0 | 0 | 0 | 0-0 | 0 | 7 / 6 |
| Maridhas Answers | 0 | 0-0 | 0 | 0 | 0-0 | 0 | 2 / 5 |
| Reflect News Tamil | 0 | 0-0 | 0 | 0 | 0-0 | 0 | 3 / 7 |
| info politics Venkee Rathinam | 0 | 0-0 | 0 | 0 | 0-0 | 0 | 7 / 6 |

"Feeling in the 8 read" is how many of the 8 minutes that draw's agents read for the channel the first reader labelled
"stirs feeling". The intervals are from a bootstrap over the agents (10,000 redraws, seed 20261123).

**What held in both draws (DEVIATION 14: readings added after the check; descriptions, not causes).** Spearman's rho
with a two-sided permutation p (10,000 shuffles):

| Label | Marks vs picks, draw 1 (written before) | Marks vs picks, draw 2 | The 8 minutes read vs picks, both draws (42) |
|---|---|---|---|
| Detail | +0.37 (p 0.104) | +0.44 (p 0.048) | +0.53 (p 0.001) |
| Questions the state government | +0.20 (p 0.385) | -0.09 (p 0.695) | +0.21 (p 0.187) |
| Questions the Union government | +0.18 (p 0.432) | +0.22 (p 0.351) | +0.09 (p 0.587) |
| Shows proof | +0.37 (p 0.100) | +0.34 (p 0.135) | +0.23 (p 0.140) |
| Makes you laugh | -0.40 (p 0.077) | -0.45 (p 0.045) | -0.51 (p 0.001) |
| Stirs feeling | -0.52 (p 0.018) | -0.54 (p 0.012) | -0.75 (p 0.000) |

- In both draws, channels whose minutes stir feeling more often were picked less. Across the 42 channel-draws, the
  more of the 8 minutes read that stirred feeling, the fewer agents picked the channel (rho -0.75); more minutes with
  a detail went with more picks (+0.53), and more that made you laugh with fewer (-0.51). The 42 are 21 channels read
  twice, not 42 separate channels.
- The same channel moved with its minutes. Vikatan TV: none of its 8 minutes in draw 1 stirred feeling, and all 99
  picked it; 5 of its 8 in draw 2 did, and none of the 20 picked it. RETROLUXE: 4 of 8 in draw 1, picked by none; 0 of
  8 in draw 2, picked by 12 of 20.
- Reading 3: six channels were picked by fewer than 5% of the agents in both draws, and in fact by no agent in either
  draw: Aadhan News, Idam Valam, Maridhas Answers, Reflect News Tamil, Arasiyal Sadhurangam and info politics Venkee
  Rathinam. On average their minutes stir
  feeling in 61.8 of 100 and give a detail in 37.5; the other fifteen channels' minutes, 43.3 and 53.2
  (`viewers-draw2/both-draws.json`, "3_mean_marks"). Not every one fits: Maridhas Answers stirs feeling in 35.1 of 100
  minutes and was picked by no agent in either draw.
- (Withdrawn 17 Nov, 08:36 IST by the file clock: an earlier line here counted "the eight channels where 60% or more of all minutes stir
  feeling" and their picks, 12 of 294 and 3 of 60. That cut at 60% was chosen after the results and is not one of
  DEVIATION 14's three readings; at 57% the same count is 25 of 294 and 6 of 60. It is left out of every claim.)
- The agents and the readers who gave the labels are the same kind of AI model (Opus). That a calm minute was picked
  and a charged one was not may be partly that model's own taste, and says nothing yet about what people choose.

**Names said aloud.** Each channel's heading became a letter; the minutes themselves were left as spoken, so some say
a name. Five channels say their own name in the minutes drawn in draw 1 (a sign-off, a greeting, a subscribe line, a
sponsor read): Vikatan TV, Behindwoods O2 (its name garbled in the captions, beside its show and host), Minnambalam,
Reflect News Tamil and Galatta Voice. What the agents wrote under "recognised" is counted two ways
(`viewers/answers-mapped.json`, `viewers-draw2/answers-mapped.json`, the "who" of each recognition): any recognition (a
channel, a show, a host or a guest; the "recognised_by" of `result.json`), and the channel's own name.
- Draw 1 (99 agents): Vikatan TV 99 by any and 99 by its name; Minnambalam 99 and 99; Reflect News Tamil 99 and 99;
  Galatta Voice 99 and 97; Behindwoods O2 99 and 0 (all 99 wrote its host's show, "Rangaraj Pandey's show", not the
  channel); Saattai 1 and 0. Of the five, the four besides Vikatan TV were picked 12, 6, 0 and 0 times, so a name
  alone did not decide.
- Draw 2 (20 agents): Savukku Media Network 20 and 20; Reflect News Tamil 20 and 20; Tribes 20 and 20; RETROLUXE 20
  and 10 (its host greets the channel's viewers by its name in minutes 2 and 6; ten agents wrote that name, as heard
  "Retrolux", and the other ten its host and show); Idam Valam 18 and 0 (a vlog's name); Saattai 18 and 0 (a show's
  name); Arasiyal Sadhurangam 18 recognitions, 15 of them naming the channel, most as "possibly"; Minnambalam 11 and 0
  (a guest); Vikatan TV 3 and 0 (a speaker); Neerthirai 1 and 0.
- **In both draws the channel on top said its own name in the minutes read** (Vikatan TV in draw 1, minute 5;
  RETROLUXE in draw 2, minutes 2 and 6), so the letters did not hide every name, and least where it mattered most.
  (Corrected 17 Nov, 08:36 IST by the file clock: this passage first said "all 99 agents named all five" and listed draw 2 by
  "recognised_by"; that count takes in an agent who named only a host, a show or a guest.)

**Names in the four later draws, and the two on top (added 17 Nov, 13:44 IST by the clock, after the six-draw result;
a count of what the answers hold, not a reading that changes the answer).** Counted as "any recognition" above (the
channel, a show, a host or a guest; `viewers-draw<d>/answers-mapped.json`, "recognised"), for the two on top and the
third. Chanakyaa was recognised by no agent in draws 1, 2, 4, 5 and 6, and was in the three of 52, 3, 11, 18 and 11
agents there; in draw 3 all 20 recognised it (a minute names the studio) and 15 picked it. Vikatan TV was recognised
by all 99 in draw 1 (its minute 5 says its name) and 99 picked it; in draw 2 by 3 agents (a speaker), none of whom
picked it; in draws 3, 4 and 6 by no agent, and it was in the three of 18, 16 and 11 there; in draw 5 by 1 agent, who
did not pick it, and 1 other picked it. Behindwoods O2 was recognised by all 99 in draw 1 (its host's show) and picked
by 12, by 1 agent in draw 3, and by none in draws 2, 4, 5 and 6, where it was in the three of 9, 17, 12 and 14. So
both channels on top were picked often in draws where no agent recognised them; a name said aloud is one weak place
of the test (the limits below), and the video says the draw-1 one ("ஒரு நிமிஷத்துல விகடன் பேரே வந்துச்சு").
(17 Nov, 14:48 IST: the posted video does not speak of the letters, so it does not say
the draw-1 name either; the names said in minutes stay a limit of the test here and on the page.)

**Four more draws, and the answer over six (DEVIATION 15, its rule written before any of them ran).** Because the
first two draws disagreed, four more draws of 8 fresh minutes a channel were read, each by 20 of the 100 profiles
drawn at random, with the same instructions, letters and order (none of these minutes in any earlier draw; minutes
that speak of a creator's arrest or legal case replaced first). A channel's share in a draw is the agents in form who
put it among their three, over the agents in form; its score is the mean of its six shares, each draw weighted the
same. The rule: the leader is settled only if it is first in at least 80% of 10,000 redraws of the six draws.

| Channel | Draw 1 (of 99) | Draw 2 (of 20) | Draw 3 (of 20) | Draw 4 (of 18) | Draw 5 (of 18) | Draw 6 (of 20) | Score | 95% interval | First in the redraws |
|---|---:|---:|---:|---:|---:|---:|---:|---|---:|
| Chanakyaa | 52 | 3 | 15 | 11 | 18 | 11 | 0.598 | 0.383-0.800 | 40.7% |
| Vikatan TV | 99 | 0 | 18 | 16 | 1 | 11 | 0.566 | 0.251-0.873 | 38.9% |
| Behindwoods O2 | 12 | 9 | 2 | 17 | 12 | 14 | 0.497 | 0.260-0.739 | 20.0% |
| AranSei | 77 | 5 | 2 | 2 | 7 | 1 | 0.280 | 0.108-0.491 | 0.4% |
| Dream Big with SK | 13 | 8 | 8 | 1 | 2 | 8 | 0.250 | 0.131-0.355 | 0.0% |
| Minnambalam | 6 | 0 | 2 | 1 | 9 | 6 | 0.169 | 0.045-0.326 | 0.0% |
| Red Pix 24x7 | 10 | 9 | 3 | 0 | 0 | 2 | 0.134 | 0.034-0.267 | 0.0% |
| IBC Tamil | 0 | 4 | 8 | 2 | 0 | 0 | 0.119 | 0.019-0.252 | 0.0% |
| RETROLUXE | 0 | 12 | 0 | 1 | 0 | 0 | 0.109 | 0.000-0.309 | 0.1% |
| Galatta Voice | 0 | 4 | 0 | 2 | 4 | 1 | 0.097 | 0.027-0.168 | 0.0% |
| Savukku Media Network | 13 | 3 | 0 | 1 | 0 | 4 | 0.089 | 0.031-0.150 | 0.0% |

The other ten channels scored 0.027 or less; Aadhan News, Idam Valam, Reflect News Tamil, Arasiyal Sadhurangam and
info politics Venkee Rathinam were in no agent's three in any draw (`viewers-six.json`, every channel and draw).

**The answer: not settled, with two channels on top.** Chanakyaa leads on score but is first in only 40.7% of the
redraws, so by the rule the result is the top two: **Chanakyaa and Vikatan TV**, with Behindwoods O2 next. Every
channel's count moved with the minutes read: Vikatan TV was in the three of 99 of 99 agents in draw 1 and of 1 of 18
in draw 5; Chanakyaa in 3 of 20 in draw 2 and 18 of 18 in draw 5. The two channels with the most political views in
the month's search (Savukku Media Network, 11.8 million; Arasiyal Sadhurangam, 8.3 million) scored 0.089 and 0.

**What the two on top do, by the marks (a description, not a cause).** Of every 100 readable minutes, Chanakyaa's
stir feeling in 8.3 and raise a laugh in 4.4, and Vikatan TV's in 28.8 and 18.6; Savukku Media Network's in 57.5 and
42.7, and Arasiyal Sadhurangam's in 59.8 and 37.5 (section 7). Behindwoods O2, third, stirs feeling in 16.9. Not every
calm channel was picked: Red Pix 24x7 stirs feeling in 17.8 of 100 minutes and scored 0.134. The agents and the label
readers are the same kind of model, so the pattern may partly be that model's own taste (the limits below).

**Party pull.** For every channel each agent also said whether it would pull them toward a party; the counts are in
`viewers/result.json` and every answer, with its reason, in `viewers/answers/` and `viewers-draw2/answers/`.

**The agents' choice beside the channels' views** (17 Nov 2026, 15:09 IST; added for the video, after the six rounds'
result, from files that existed before any agent read: no part of the test changed). Two measures of views, among
the 21:
- *This month's political search views* (`channels.json`, sweep_political_views; the sweep that chose the sixteen):
  1. Savukku Media Network 11,786,468;
  2. Arasiyal Sadhurangam 8,305,517;
  3. Chanakyaa 5,296,890;
  4. Saattai 4,005,776.

  Vikatan TV is 14th.
- *The listed views of every upload of the two months' window* (`kept.json`, by_channel, window_listed_views, long
  videos and Shorts together, political or not, as the channel listing shows them; live streams not counted; a range
  where the window's edges are not fixed by the listing's order and the dates read):
  - Saattai 74.1–79.1 million: its least is above every other channel's most.
  - Savukku Media Network 59.6–66.2 million and Chanakyaa 56.5–67.2 million: 2nd and 3rd in some order.
  - Vikatan TV 24.9–27.0 million: 8th to 10th.
- *A limit on the most.* Six channels' Shorts listings (the newest 600, or 800 for Chanakyaa) may end inside the
  window, so their most may be short by the Shorts beyond the listing:
  - Behindwoods O2, Chanakyaa, Galatta Voice, IBC Tamil, Minnambalam and Vikatan TV.
  - Chanakyaa's Short at position 706 is dated 24 Sep 2026 on its own page (`uploads.jsonl`), so the window starts
    near the listing's end at 800.
  - To pass Saattai's least, Chanakyaa would need about 275 more Shorts in the window at its last hundred's average
    of 25,547 views.
  - The other five are at 42.1 million or less.
- *The agents' six-round score*:
  - for the two first by views: Savukku Media Network 0.089 (21 of the 195 answers in form) and Saattai 0.008 (1 of
    195);
  - for the agents' two: Chanakyaa, top 3 by both measures, 0.598; Vikatan TV 0.566.
- *A description, not a cause.* The agents' two stir feeling in fewer minutes (8.3, 28.8 of 100) than the views' two
  (57.5, 74.4). Chanakyaa, the calmest of the 21, is also in the top three by both measures of views. Views count
  watching, not people who follow, and the agents read only the words, with no face, voice or fandom.

**The second measure, read deeper: it corrects the note above** (17 Nov 2026, 16:01 IST; DEVIATION 16 and its result in
DEVIATIONS.md; `window-views-deep.json`). A fresh review of the video found that the six Shorts listings' "most" was
not a bound. They were read again deeper, by a rule written first, and each now ends before 17 Sep 2026.
- *The listed views of every upload of the two months*, all 21:
  1. Galatta Voice 95.46–95.51 million (4,255–4,263 Shorts in the window, about 70 a day);
  2. Saattai 74.1–79.1 million;
  3. and 4. Chanakyaa 65.1–67.4 million and Savukku Media Network 59.6–66.2 million, in some order.
  - Then Behindwoods O2 46.2–46.6, IBC Tamil 44.5–44.6, Aadhan News 33.9–42.1, Minnambalam 38.9–40.1; Vikatan TV
    25.1–27.4 million, 9th to 11th.
- So "Saattai first by the two months' uploads" in the note above is wrong: it is second, after Galatta Voice. The
  an earlier reading said it; the video says the order above.
- Chanakyaa is 3rd by the search views (counting its collaboration label, by DEVIATION 2; 4th by its own views alone)
  and 3rd or 4th by the uploads. Galatta Voice's six-round score is 0.097; it is 19th by the search views.
- *Feeling against views, over the 21* (Spearman's rank correlation, permutation p): with the agents' six-round score
  −0.82 (p < 0.001); with the search views +0.12 (p 0.60); with every upload of the two months −0.11 to −0.08 (p 0.65
  to 0.73); with long videos alone −0.14 (p 0.56). The agents' lean to calm channels is strong; views show no pattern
  either way. The description above stands as a description of the four channels, not of views.

**Corrections after a later review** (17 Nov 2026, 16:27 IST; a fresh review, and one finding of
the check's own, after the deeper listings).
- *What the 15:09 note above withdraws* (the 16:01 note gives the order that replaces them): "Saattai first by the
  two months' uploads" (second, after Galatta Voice); "Savukku Media Network and Chanakyaa 2nd and 3rd in some order"
  (3rd and 4th); "Vikatan TV 8th to 10th" (9th to 11th); "the other five are at 42.1 million or less" (Galatta Voice is
  at 95.5 million); "about 275 more Shorts" (Galatta Voice alone has 4,255 to 4,263 in the window); and "Chanakyaa, top
  3 by both measures", said twice (3rd by the search views, counting its collaboration label; 3rd or 4th by the
  uploads, not sure which).
- *Live streams.* The second measure sums the Videos and Shorts tabs only; live streams are not counted (DEVIATION 1),
  political or not, and no file holds their views, so whether the order holds with them is not known. The video says
  "எல்லா வீடியோஸ்" in its voice; its screen says "அரசியலும், மத்ததும் · live தவிர".
- *The share of the window's views* (section 6): corrected beside the table and in the two lines that gave it; all 21
  together hold 18.8% to 20.2%, not 22% to 25%.
- *Three channels' Shorts were ranked within the newest part of the window only* (the check's own reading of the
  deeper listings; its counts corrected at 16:43 IST after a fresh check, which counted by the Shorts' ids where the first
  version counted by positions, and 45 Shorts were put up between the two reads). The kept Shorts were ranked by views
  within the first listing (600 Shorts each). For Galatta Voice and IBC Tamil that listing reaches back only to about
  5 Nov 2026; for Behindwoods O2 to somewhere between 22 Sep and 9 Oct (its last dated Short is from 9 Oct). Matched
  by id, the deeper listings show 3,655 to 3,660, 2,837 to 2,839 and 147 to 152 Shorts of the window that the first
  listing did not hold; in the deeper read, 115, 64 and 8 of them have more listed views than the least-viewed of the
  channel's five kept Shorts in the same read. Their titles were not read, so whether any is political is not known.
  So for these three, "the most-viewed political uploads of the window" is shown for the long videos and not shown for
  the Shorts: their five Shorts are the most-viewed political Shorts of the newest 600. For
  Chanakyaa, Minnambalam and Vikatan TV the first listing reached every window Short with more views than their kept
  ones. The marks and the agents' minutes are unchanged; this is a limit on how far the Shorts stand for these three
  channels.
- *The later read.* The Shorts read in both listings gained 0.4% (Behindwoods O2) to 6.9% (IBC Tamil) in the about 11
  hours between them (Minnambalam 0.7%, Chanakyaa 0.7%, Galatta Voice 1.6%, Vikatan TV 2.1%). DEVIATION 16's "about
  16 hours" was its estimate before the read.
- *The page probes were not paced as the tool says.* Each probe was a fresh process, so the pace inside it never
  applied between probes: 76 of the 79 gaps between the 80 pages read for the edges are under 3 seconds (09:59 to 10:08
  UTC). YouTube refused none. A probe that could not be read leaves no file, so the one unread page at Vikatan TV's
  start edge (its range 527 to 626 Shorts) cannot be shown from the files.
- *What is published of the deeper read.* The deep listings are published without their title column, and the pages
  with their dates and counts only, as DEVIATION 10 requires of raw listings (1,068 of the listed titles carry a
  police or court word and were never read for a creator's arrest or case).
  *Added 16:43 IST:* the published deeper read is `window-views-deep.json`, `tools/window_edges.py`, `listings-deep/` and
  `pages-deep/`; for the six Shorts tabs it replaces the window sums and shares in `kept.json`, which are the first
  listings' and are left as they were read.

## The test's limits
This is one test of a fixed size, and its answer holds only for what it read.
- **Its size.** 295 videos (210 long videos and 85 Shorts) from 21 channels, 5,763 minutes, of
  which 5,752 had text and were read. They are each channel's most-viewed political uploads of the window, and
  they hold between 22% and 25% of the views of the channels' uploads in the window (per channel in the table below).
  *Corrected 17 Nov 2026, 16:27 IST:* between 18.8% and 20.2% with the deeper Shorts listings (DEVIATION 16), and
  for three channels the Shorts are the most-viewed political Shorts of the newest part of the window only (section 8,
  "Corrections after a later review").
- **What it does not cover:**
  - live streams;
  - videos older than 17 Sep 2026, and the less-viewed political videos of the window;
  - comments;
  - the real recommendation feed in India: no agent browsed, and nothing here measures what YouTube actually shows
    a newcomer in Tamil Nadu.
- **The readers are machines.** The minutes' text is YouTube's machine captions, with misheard words. The labels
  are AI readings, and two AI readers agree on each label only as far as the kappa above says. The viewer agents are
  AI agents, not people.
- **8 minutes a channel, and the answer moved with them.** Each agent read 8 minutes of each channel, from 1.3% (Savukku Media Network, 602
  readable minutes) to 13.8% (Reflect News Tamil, 58) of a channel's readable minutes. A second draw of 8 minutes, read by 20 agents, put a different channel on top
  and the first leader at none of 20; over six draws the leader was first in only 40.7% of the redraws (section 8).
  Which channel leads is not settled by this test: two channels are on top.
- **The agents and the readers are one kind of model.** The viewer agents and the label readers are both Opus, so a
  pattern between the labels and the picks may partly be that model's own taste.
- **The same experiment with more data and more compute may give a different answer.** More videos a channel,
  older videos, live streams, more readers on each minute, or more agents could each move the marks and the
  agents' choices.

## Every change to the plan
Each is in `DEVIATIONS.md` with its time and its full reason. In short:
1. **Videos chosen by views, not all read** (to finish in about six hours): each channel's most-viewed political
   uploads of the window, 10 long and 5 Shorts; every minute of them read by AI readers, so no model was trained.
2. **21 channels, not 20:** the plan's own rule added Reflect News Tamil, which its written list had missed.
3. **After YouTube refused the page reads:** political or not read in two stages (title first, then title and
   description), one visit a page, and the window's views given as a range where the window's edges were not found
   exactly.
4. **The viewer profiles' figures:** Census 2011 for age, gender and place, and a 2021 post-poll survey for issues,
   because the 2026 roll by age and a 2026 survey with issue shares could not be found.
5. **Pages read again before the hour's back-off was up,** after one probe passed, at one page every 10 to 11 seconds.
6. **A mark is a share of the readable minutes,** with the shares of minutes without text and of unreadable minutes
   beside it; minutes labelled in batches as videos were fixed; page visits replanned every 30 pages.
7. **A last time for page visits** (06:45 IST), the tabs not filled by then to keep what they had. The visits ended
   at 06:42 IST with nothing left to open, so it cut no tab.
8. **The agents' 8 minutes a channel come from readable minutes only.**
9. **Two of the agents' minutes were replaced** because they speak of a creator's arrest or legal case, which this
   round keeps out of every output; each by another readable minute of the same channel. Labels and marks unchanged.
10. **Nineteen titles withheld** in the published files, because they may speak of a creator's arrest or legal case;
   the uploads and every count stay. The raw listings are to be published without their title column.
11. **The agents ran in waves of 10 with a budget rule;** all 100 ran (the week's meter went from 69% to 76%).
12. **One answer out of form left out** (P069, "toward TVK" for "TVK"), as the rule written before any answer says;
   the result is on 99. With it counted, the leader is the same.
13. **A check added after the result:** 20 agents read a fresh 8 minutes of each channel, because every agent in draw 1
   read the same minutes. Its rule was written before it ran; it changed the leader, so the leader is said as not
   settled.
14. **Readings added after the check,** to say what held across both draws; said as added after, and published whole.
15. **Four more draws of minutes, and one rule for the answer over all six,** written before any of them ran: 80 more
   agent answers (76 in form), minutes about a creator's arrest or case replaced first; the result is not settled,
   and the top two are said (the week's meter went from 85% to 91%).
16. **The Shorts listings read deeper** (added 16:27 IST): six Shorts tabs whose first listing could end inside the window were
    listed again, past 17 Sep 2026, and the window's edges found by the Shorts' own pages; the second measure of views
    is taken from them (Galatta Voice first, Saattai second).

## What is published with this record
- `PLAN.md`: the method as fixed before any data, and `DEVIATIONS.md`: every change, with its time and why.
- `channels.json`: the 21 channels and the rule that chose them.
- `listings/`: each channel's listed uploads, with their positions, views and the time of listing (published without
  the title column; DEVIATION 10).
- `uploads.jsonl`: every upload gone through, in view order, with its date, length, the title readers' answers, the
  political readers' answers and reasons, and what happened to it (19 titles withheld; DEVIATION 10).
- `kept.json`: the kept videos, and for each channel how far down its ranking the reading went and what share of its
  window views the kept videos hold.
- `captions.json`: each kept video's caption status; `windows.jsonl`: every minute's labels, the first reading and,
  for 300 minutes, the second.
- `marks.json`: the marks with their intervals, the views-weighted column, the share of minutes without text, and the
  agreement.
- `profiles.md`, `viewer-instructions.md`, `excerpts/`: the 100 profiles, what each agent is asked, and the minutes
  they read (with the key to the letters).
- `summary.json`: the counts in section 6, made from the files above.
- `viewers/answers/` and `viewers-draw2/answers/`: all 120 answers as the agents wrote them (P069's too);
  `viewers/result.json`, `viewers-draw2/result.json` (the tallies), `answers-mapped.json` (each answer with its letters
  mapped back to channels), `viewers/marks-vs-picks.json`, `viewers-draw2/both-draws.json`, `viewers-draw2/agents.json`
  (the 20 agents of the check) and `excerpts-draw2/` (the check's minutes, without the 5 that speak of a creator's
  case). For draws 3 to 6: `viewers-draw<d>/` (agents, answers, tallies) and `excerpts-draw<d>/` (the minutes read,
  the word hits and the readers' screens, the replaced ids without their text), and `viewers-six.json` (the answer
  over six draws, by `tools/viewers_six.py`). The packets the agents read are not
  published as files: `tools/packets.py` rebuilds every one of them, byte for byte, from `profiles.md` and the
  excerpts.
- The instructions every AI reader followed (`instructions-*.md`) and the code (`tools/`).

Kept beside the record and not published with it: `windows-text.jsonl`, the full caption text of every kept minute
(the creators' own words as YouTube's machine heard them), kept so that every label can be checked against its text;
`excerpts-excluded.json`, the ids of the two replaced minutes (DEVIATION 9); and `titles-withheld.json` (DEVIATION 10).
The 168 minutes the agents read are published in `excerpts/`. One sentence of one agent's answer in draw 2 (P014, under
"recognised") says what a creator said of his own time in jail, from one of the 5 minutes withheld there; it is
withheld in the published answer and in `answers-mapped.json`, marked where it stood, and the rest of the answer is
as written (`tools/publish_pack.py`, which finds it by its sha256; added 17 Nov, 13:49 IST by the clock).
