# Deviations from PLAN.md

PLAN.md (sha256 693d4ad196876039771ad83af44ad67f80e6e4e29ef6313432471631349e8720) is kept as fixed, except for the lines reworded for publication that README.md lists (the file as fixed has the sha256 given here; the published file's differs for that reason). Every change to
it is written here with its time (IST), what changed and why, before the result it could touch is seen wherever that
is possible.

## DEVIATION 1

DEVIATION 1 (17 Nov ~04:50 IST, before any caption was read or any label drawn): the best data is chosen by views, not every video read, so that the round finishes in about 6 hours. Replaces PLAN.md parts 2 and 5 and the live table. The videos: for each of the 20 channels, its uploads from 17 Sep to 16 Nov 2026 (long videos and Shorts; live streams out, said as out) ranked by view count as listed at selection time (time recorded). Going down that list, the first 10 long videos and the first 5 Shorts that two Opus readers find political (title and description; a third reader on disagreement) are kept. The reason it is by views: the question is which channel pulls a newcomer, and what a channel's audience actually meets is mostly its most-viewed videos; the share of the channel's window views that the kept videos hold is published beside its marks. Every kept video's every minute is read by Opus (no classifier: with this many minutes, Opus reads them all). A second Opus reader reads 300 windows drawn at random with seed 20261118, blind to the first; kappa per label is published. The marks are shares of minutes, with an interval from a bootstrap over videos, and a second column weighting each video by its views. The viewer agents' 8 minutes a channel are drawn with seed 20261120 from the kept videos.

(Entered in this file at 04:29 IST by this machine's clock, before any listing of this round was read.)

## DEVIATION 2 (17 Nov 2026, 04:33 IST, before any video of the channel was listed or read)
**What:** the channel list has 21 channels, not 20. Reflect News Tamil is added.
**Why:** PLAN.md part 1's rule counts a collaboration label with its own channel. Each collaboration label's uploader
was read from the video's own page (17 Nov, 04:30 to 04:40 IST):
- "Chanakyaa and Jayashree Tarot Reader & astrologer" (CcOtSXfFvRs, 1,449,436 views) is Chanakyaa's: Chanakyaa was
  already in.
- "VIPRA TV and 2 more" (YmkfT14ygBw, 963,757) is VIPRA TV's: under 1,000,000 with no other video, so out.
- "Reflect News Tamil and Reflect Voice" (1MGMxJrxZdc, 109,238) is Reflect News Tamil's. Reflect News Tamil had
  931,671 of its own, so 1,040,909 in all: over 1,000,000, kind interview_platform, so in by the rule.
- The others (Reflect Lifestyle, Nakkheeran TV, kumudam, Veeraa Media, Wow Tamizhaa Pulse, Tamizhian pulse, Fine Time
  Media, and "YouTamil Oodaruppu and You Tamil LIVE" at 160 views) stay under 1,000,000 with their collaboration added.
The plan's written list of 20 missed this. The rule is the method, so the rule is followed and the list corrected.

## DEVIATION 3 (17 Nov 2026, 04:42 IST, before any video was read as political or not and before any caption was read)
**What happened:** at 04:41 IST YouTube refused this machine's page reads (HTTP 429) after 218 video pages (12 of them
read from 04:30 to check the channels and the collaborations, the rest from 04:39 to find the window's edges) and 36
channel listings in about 12 minutes, with up to six paced streams at once: too many. Every reader stopped at once. As this
round's rules say, the round backs off from YouTube for an hour (to 05:41 IST) and then resumes, one paced stream at a
time, at most two.
**What changes, to finish in the 6 hours of DEVIATION 1 with far fewer page reads:**
1. **The political reading is done in two stages.** First, two Opus readers read the TITLE of every upload high in a
   channel's view ranking (the top 60 long videos and top 30 Shorts by listed views; more if a channel runs short),
   each answering political, not political or cannot tell (`instructions-title-screen.md`). Second, every upload that
   is not "not political" for BOTH readers has its own page opened, and two fresh Opus readers read its title and full
   description (`instructions-political.md`, a third on a disagreement). Only that second reading keeps a video. An
   upload that both title readers find clearly not political is not opened, and its description is not read: that is
   said beside it in `uploads.jsonl`.
2. **One page visit per opened video.** The page's own date, length, live flag and description are read in the same
   browser visit as its transcript (`tools/transcripts.mjs`), so a kept video is never fetched twice.
3. **The window's edges.** Each candidate's date is the exact one on its own page, so no video outside the window is
   kept. The window's edges were found exactly (by halving each tab with exact dates) for 12 channels' long videos
   before the refusal. For the other tabs, the window's total views (the denominator of "share of the channel's window
   views") are given as a range: from the uploads that are surely in by the listing's own dates and by the dates already
   read, to those that may be in. Which tabs have an exact count and which a range is published.

## DEVIATION 4 (17 Nov 2026, 04:56 IST): the profiles' figures, before any viewer agent read

PLAN.md part 6 draws the 100 profiles from "Tamil Nadu's adult population by age and gender, Census and the 2026
roll" and "issues voters named, the 2026 post-poll surveys". Two of those figures could not be found, and the nearest
published ones are used instead, as said in `profiles.md`:
1. **Age and gender: the Census 2011 table (C-13), not the 2026 roll.** The roll's electors by age band and gender
   could not be read (the Chief Electoral Officer's site refused the connection on 17 Nov 2026; the news reports of
   the final roll of 23 Feb 2026 give only totals). The Census is fifteen years older than the roll, so today's
   adults are on the whole older than its shares say.
2. **What they want: Lokniti-CSDS's Tamil Nadu post-poll survey of 2021, not a 2026 one.** No 2026 post-poll survey
   with issue shares was found. The 2021 survey is the same state and the same kind of survey, one election earlier;
   a few of its answers are about events of that time. Answers naming a party, leader or candidate (554) and "some
   other issue" (59) are left out, because the newcomers lean to no party and the latter names nothing; so 54 of the
   100 name no one issue.
The profiles were fixed (tools/profiles.py, seed 20261121, largest remainder; a rerun gives the same table) before any
viewer agent was started.

## DEVIATION 5 (17 Nov 2026, 05:23 IST): YouTube read again before the hour was up, slower, after one probe
DEVIATION 3 set the back-off to 05:41 IST. At 05:17 IST one plain request to a watch page was answered normally
(HTTP 200), so at 05:18 IST one paced probe was made with the visit reader itself (one page, its captions read); it
passed. Reading resumed then, before 05:41, to keep to the round's time, at a slower pace than the one that drew the
refusal (218 pages in about 12 minutes, up to six streams, about one page every 3 seconds): **one stream, about one
page every 10 to 11 seconds**, and no pictures, video or fonts fetched with a page (only the page and its text). Five
pages were read from 05:18 to 05:23 IST while the pace was measured. Any refusal stops every reader at once, and the
back-off of an hour starts again.

## DEVIATION 6 (17 Nov 2026, 05:40 IST, before any minute of this round was labelled)
1. **A mark's denominator is the channel's readable minutes.** PLAN.md part 4 makes a mark "the share of its political
   minutes where the label is true". Some kept videos have captions that are not Tamil speech at all (YouTube's English
   or Hindi recogniser run on Tamil speech: a few stray words for a whole video) or only music. A reader marks such a
   minute "not readable", and every label of it false. Counting those minutes as "label false" would lower the marks of
   channels with poor captions for a reason that has nothing to do with what they say. So a mark is the share of the
   channel's minutes **with text that the reader found readable**; the shares of minutes without text and of minutes
   found unreadable are published beside it, for every channel.
2. **The minutes are labelled in batches as the videos are fixed.** To keep to the round's time, the minutes of videos
   that no later reading can push out of a channel's kept set are labelled while the page visits go on; the rest
   after. Each batch is shuffled (seed 20261117 plus the batch's first packet number), so every packet still mixes
   channels. The 300 minutes for the second reader are drawn (seed 20261118) only after every kept video is fixed.
3. **The page visits replan every 30 pages.** Before each round of 30 visits, the list of uploads still needed is made
   again from what is known (in each tab not yet filled, in view order, the uploads still to open within the quota
   plus one), so pages that are no longer needed are not opened. Same pace, one stream.

## DEVIATION 7 (17 Nov 2026, 05:48 IST, before the kept set is fixed): a last time for page visits
At one page every 10 to 11 seconds (DEVIATION 5), the page visits are the slowest part of the round. To keep to the
round's time, **page visits stop at 06:45 IST at the latest**. A tab that has not reached its 10 long videos or 5
Shorts by then keeps what it has; how many it has, and how far down its view ranking the reading went, is published
for every tab (`kept.json`, "by_channel"), and said in the record. Nothing else changes: the order of reading stays
by views.

## DEVIATION 8 (17 Nov 2026, 05:58 IST, before any minute is drawn for the viewer agents): the agents' minutes are drawn from readable minutes
The plan draws each channel's 8 minutes for the viewer agents from its minutes with text. Some automatic caption tracks
are labelled with another language and hold misheard text in another script, not Tamil (DEVIATION 6). A minute like
that would show the agents a channel as noise. So the 8 minutes are drawn, with the same seed (20261120), only from the
minutes the first label reader found readable. How many of each channel's minutes were unreadable is published beside
its marks.

### What DEVIATION 7 came to (17 Nov 2026, 06:44 IST, written after the kept set was fixed)
The page visits ended at 06:42:54 IST, before the 06:45 IST limit, because no upload was left to open: **no tab was cut
by the limit.** Each of the 42 tabs reached its 10 long videos or 5 Shorts, except four Shorts tabs (Idam Valam,
Maridhas Answers, Red Pix 24x7, info politics Venkee Rathinam) whose newest Short 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 (`kept.json`, "by_channel": `not_reached` is 0 everywhere).

One fault was found and mended before that: the page-visit planner took a tab to be full as soon as 10 of its uploads
were marked kept, even when uploads ranked above the tenth had not been opened. In Idam Valam's long tab, ranks 10, 11
and 12 by views had not been opened while rank 13 was marked kept. They were opened at 06:42 IST and read by two
readers (all three political, in the window), so ranks 1 to 10 are kept and rank 13 is not. The planner now treats a
tab as full only when nothing above its last kept upload is undecided (`tools/keep.py`, `cmd_visits`). No other tab
had an undecided upload above a kept one. The minutes labelled before this were only from videos no later reading
could push out, so no labelled minute was affected.

## DEVIATION 9 (17 Nov 2026, 06:54 IST, before any agent has read anything): minutes about a creator's arrest or case are replaced
The round's rules keep the arrests and legal cases of creators (the people who run or speak on the channels) out of
every output, and the agents' minutes are published. The 168 drawn minutes were searched for police and court words
(43 minutes hold one), and an AI reader (Opus) read those 43 for a creator's arrest or case: 2 minutes speak of one.
Each is replaced by another readable minute of the same channel, drawn with the seed "20261120-<channel key>" from the
channel's minutes not yet drawn, so no other channel's minutes change; a replacement is checked the same way. The ids
of the replaced minutes are kept beside the record (`excerpts-excluded.json`), not published. The labels and marks are
not changed: every minute is still read and counted; only what is shown and published as text changes.

## DEVIATION 10 (17 Nov 2026, 06:55 IST, before anything is published): titles that may speak of a creator's arrest or case are withheld
The published files `uploads.jsonl` and `kept.json` carry YouTube's own titles of the 1,914 uploads gone through, and
the readers' one-line reasons. The titles and reasons were searched for police and court words (131 uploads hold
one), and an AI reader (Opus) read those 131 for a creator's arrest or legal case, answering yes where it could not
tell: 19 uploads (2 of them kept videos). Their titles and reasons are replaced by a note in both files; the uploads,
their ids, their states and every count stay as they were, and the measurement is unchanged. The raw listings
(`listings/`, 28,024 listed uploads as YouTube listed them) were not read this way: they are kept as fetched, and
are to be published without their title column unless they are read the same way first.

## DEVIATION 11 (17 Nov 06:58 IST, before any viewer agent ran): the run's budget
The week's plan stands at 69% with four days to its reset, and the cost of 100 agents each reading about 184,000
characters is uncertain (two estimates: about 11% and about 35% of a week). So the agents run in waves of 10, in the
order P001, P002, ... (the profiles' order is their draw order, not sorted by any trait). After the first wave the cost
of one agent is measured from the plan's meter, never from any answer, and the run goes on to all 100 if the meter
after all 100 is projected at 88% or under; otherwise it stops at the largest whole wave that keeps it there, and never
under 50 agents. No answer is opened before the run ends. The number that ran is said wherever the result is said.
The reading is exactly the planned one: each agent's packet (viewers/packets/P###.md) is viewer-instructions.md's prompt
filled with its profile, then the 21 channel files in its own order under its own letters (excerpts/agents.json), lines
wrapped under 1,500 characters at a space so no reader truncates a line (no word changed). Each answer is written whole
to viewers/answers/P###.json.

DEVIATION 11, measured (07:01 IST, before any answer was read; only the schema check was run, which prints pass or
fail): wave 1 (P001-P010) all passed the schema. The plan's meter went 69% -> 70% across the wave, and the session's
API-priced cost line rose about $18 (1258.6 -> 1276.7), so 100 agents are projected at about $180, about 7-8% of a
week, about 78% on the meter: under 88%, so all 100 run (P011-P100 in waves of 30).

## DEVIATION 12 (17 Nov 07:11 IST, before the tally was run): one answer out of form
All 100 viewer agents answered. 99 pass the schema. P069 wrote "toward TVK" (not "TVK") in four of its 21 party
answers. tools/viewers_tally.py, written before any answer, leaves an answer out of form and mends nothing by hand,
so the result is on 99 agents, as that rule says. Its answer file is kept whole and published. After the main tally,
the same tally is run once more with only that prefix removed (a check, said beside the result: whether the
leader changes with it).

## DEVIATION 13 (17 Nov 07:13 IST by the file's write time, after the main result was seen; written before the check runs)
**What was seen:** the main result (viewers/result.json, 99 agents): Vikatan TV in 99 agents' three, first for 93.
Every agent read the same 8 minutes of each channel, so the bootstrap interval is over agents only, never over which
minutes were drawn. Vikatan TV's 8 drawn minutes have no minute labelled "stirs feeling", while 29% of all its
readable minutes are (marks.json). Its minutes also say its name aloud, and all 99 agents recognised it (four other
channels were recognised by all 99 too, and were picked by 12, 6, 0 and 0).

**The check added, and why:** a result that could flip with another draw of minutes must be said as not settled
(research-integrity). So one check with a second draw:
- **Minutes:** for each channel, 8 minutes drawn with seed 20261125 from its readable minutes that were not in the
  first draw and not in excerpts-excluded.json. Built by tools/excerpts_draw2.py into excerpts-draw2/.
- **Agents:** 20, drawn at random from the 100 profiles with seed 20261126. Each keeps its own letters and order from
  excerpts/agents.json and the same instructions; only the minutes change. Packets are written to
  viewers-draw2/packets/, answers to viewers-draw2/answers/.
- **Tally:** the same rules as tools/viewers_tally.py (top-three counts, first choices, a bootstrap over the 20
  agents), by tools/viewers_tally_draw2.py.
- **What it decides:** it is reported beside the main result and never replaces it.
  - If Vikatan TV is also the most picked in the check (alone or tied), the video and the page say the leader held on a
    second set of minutes read by 20 agents.
  - If not, they say the leader is not settled: it depends on which minutes were read.
- **Cost:** about $40, measured on the meter before and after.

Note to DEVIATION 13 (07:16 IST): its heading first said 07:16 and tools/viewers_tally_draw2.py first said 07:22;
both times were written ahead of the clock. Corrected to the files' write times: DEVIATION 13, the draw and the packets
07:13-07:14 IST; the check's tally script 07:15 IST, when no check answer existed (viewers-draw2/answers/ was empty).

## DEVIATION 14 (17 Nov 07:19 IST, after the check's result; written before these readings are run)
**The check's result** (viewers-draw2/result.json, 20 of 20 in form): RETROLUXE 12 of 20, Behindwoods O2 9, Red Pix
24x7 9, Dream Big with SK 8, AranSei 5, ... Vikatan TV 0 of 20. By DEVIATION 13's rule the leader is **not settled**:
which channel the agents pick depends on which 8 minutes they read. The main result stands as measured and is
published, always with the check beside it.

**Readings added, after the fact, to say what held across both draws** (descriptions, not causes; said as added after
the check), by tools/viewers_both.py:
1. The pre-written description (tools/viewers_marks.py) run on the check too: Spearman between each channel's picks
   and its six marks (all its minutes), permutation p (10,000 shuffles, seed 20261124).
2. The minutes the agents actually read: for each channel and draw (21 x 2 = 42), the share of that draw's agents
   who picked it among their three, against how many of its 8 drawn minutes carry each label (the first reader's
   labels in windows.jsonl). Spearman, permutation p (10,000 shuffles, seed 20261127).
3. The channels picked by fewer than 5% of the agents in both draws, with their marks.
Nothing is chosen from these by its p-value; all three are published whole.

Note to DEVIATION 11 (07:20 IST, found when the packet builder was written out as tools/packets.py): the prompt
as sent differs in wording from viewer-instructions.md, never in what is asked. The exact text sent is HEAD in
tools/packets.py, and it rebuilds all 120 packets byte for byte. The four differences:
1. The profile is said as one sentence ("a 35-year-old woman living in a city in Tamil Nadu. What you want from a
   political video: ...") in place of the template's "{profile line}. You live in Tamil Nadu."
2. Question 2 adds how to write it: (written "away from <party code>").
3. The answer's shape is given on one line, with the agent's own id and the letters it was shown, in place of the
   template's longer example; the schema rules are the same.
4. A line ends the prompt: "The channels follow, in the order you read them."

## Note on times (07:53 IST): what the clock and the files say
Several times in PLAN.md and in the headings of DEVIATIONS 1 and 2 were written by estimate, not read from a clock.
PLAN.md is kept as fixed, so they are corrected here. Each true time comes from the files' own write times on this machine and from this record's own entries.
- 04:24:49: PLAN.md written (its write time; the sha256 above recorded at 04:25 in the reel's progress file). PLAN.md's
  first line says "about 04:40".
- 04:27: the choice to select by views, not read every video, was made. DEVIATION 1's text says "~04:50"; it was
  entered at 04:29 by this machine's clock, as its own last line says.
- 04:44:51: channels.json written, with the 21 channels. DEVIATION 2's heading says 04:33, while its text says pages
  were read from 04:30 to 04:40, so its heading time is not the time it was written. That it came before any listing
  holds by the files: the first listing was written at 04:47:42.
The order the method depends on holds: the choices of the method, then the plan, then the change to "by views", then
the channels, then the first listing, then every reading.

## DEVIATION 15 (17 Nov 12:50 IST by the clock; written before any minute of these draws is drawn or read)
**Why:** the check changed the leader, so the two draws disagree. Four more draws of minutes are run, then one winner
by a rule fixed here, before any of them runs. Whichever channel the six draws name is the answer, and the video's
justification follows it.

**The four draws (draws 3, 4, 5 and 6), run in that order:**
- **Minutes:** for each channel, 8 minutes drawn with seed 2026113d (draw 3: 20261133; 4: 20261134; 5: 20261135;
  6: 20261136) from its readable minutes that are in no earlier draw (1, 2, and the earlier of 3 to 6) and not in
  excerpts-excluded.json. Same channel files, letters and order as the first two draws. Built by
  tools/excerpts_drawN.py into excerpts-draw<d>/.
- **A creator's arrest or case** (DEVIATION 9): the drawn minutes are searched for police and court words (the list is
  in tools/excerpts_drawN.py); an AI reader (Opus) reads every minute that holds one, and a minute that speaks of a
  creator's arrest or legal case is replaced by another readable minute of the same channel not yet drawn (seed
  "2026113d-<channel key>"), checked the same way. Draw 2's minutes, which were not screened, are screened now for
  publication only (its answers stand as read).
- **Agents:** 20 a draw, drawn at random from the 100 profiles with seed 2026114d (draw 3: 20261143 ... draw 6:
  20261146), each draw on its own (a profile may sit in more than one draw). Each keeps its own letters and order from
  excerpts/agents.json and the exact prompt of tools/packets.py; only the minutes change. Packets to
  viewers-draw<d>/packets/, answers to viewers-draw<d>/answers/. Opus agents, one per packet.
- **Out of form:** an answer that fails the schema is left out and listed (as DEVIATION 12); nothing is mended or rerun.

**The winner, decided by this rule and no other:**
- For each channel and draw, its share = agents in form who put it among their three / agents in form in that draw.
- Its score = the mean of its shares over all six draws (1 to 6), each draw weighted the same (draw 1 has 99 agents,
  the others 20: equal weights stop one draw of minutes from deciding).
- The leader = the highest score. **Settled** only if the leader is first (ties shared) in at least 80% of 10,000
  bootstrap redraws over the six draws (six draws drawn with replacement; seed 20261150). If not, the result is "not
  settled", and the top two by score are said with their share of redraws first.
- Also published beside it: each draw's top three; first choices; the leader's score with a 95% interval from the same
  redraws. Computed by tools/viewers_six.py, written before any answer of draws 3 to 6 exists.
- **What it decides:** the video names the settled leader as the channel these AI agents picked most over six draws of
  minutes, or says "not settled" with the top two. The justification is what its minutes do, by the marks
  (marks.json), said as a description and not a cause. Whatever comes out stands.
- **Cost:** about $2 an agent, 80 agents; the meter is read before and after.

### How DEVIATION 15 ran (written 17 Nov, 13:09 IST by the clock, after all four draws were tallied)
- **The screens.** Police or court words were found in 44 of draw 3's 168 minutes, 45 of draw 4's, 39 of draw 5's
  and 44 of draw 6's. The Opus reader judged 5, 8, 1 and 2 of them to speak of a creator's arrest or legal case;
  each was replaced as the rule says, and every replacement was searched and read the same way. In draw 4 two
  replacements were themselves judged unclear (whether the person spoken of was a creator) and were replaced again,
  so draw 4 replaced 10 minutes in all; one replacement in draw 4 and one in draw 6 spoke of politicians' cases and
  were kept. The ids are in excerpts-draw<d>/replaced.json (5, 10, 1 and 2), the readers' answers in screen.json
  and screen-replacements.json. Draw 2's minutes, screened for publication only: 46 with a word, 5 judged to speak of
  a creator's case; those 5 are withheld from the published excerpts of draw 2, and its answers stand as read.
- **Out of form**, left out and not rerun: draw 4, P037 and P078; draw 5, P009 and P069 (each wrote "toward TVK"
  where the form asks "TVK"). Draws 3 and 6: none. Agents in form: 99, 20, 20, 18, 18, 20.
- **The result, by the rule above and no other** (tools/viewers_six.py, viewers-six.json): **not settled.** The
  leader, Chanakyaa (score 0.598, 95% interval 0.383 to 0.800), is first in 40.7% of the 10,000 redraws, short of 80%.
  The top two: Chanakyaa 0.598 (first in 40.7%) and Vikatan TV 0.566 (first in 38.9%). Next: Behindwoods O2 0.497
  (first in 20.0%); every other channel 0.280 or less.
- **Cost:** the week's meter read 85% before the four draws and 91% after (data/usage/plan.jsonl).

## DEVIATION 16 (17 Nov 2026 15:26 IST by the clock; written before any deeper listing is read)

**What.** The video compares the agents' choice with two measures of the channels' views. One is the
listed views of every upload in the two months' window (`kept.json`, by_channel, window_listed_views). A fresh
reviewer found that for some Shorts tabs that sum is not bounded above. The tab's listing (the newest 600, or 800 for
Chanakyaa) may end inside the window. For Galatta Voice and IBC Tamil the oldest dated Short in the listing is from
5 Nov 2026, so most of their window's Shorts were never listed.

**Why.** The measure claims "every upload in the window". Incomplete listings make it a different measure for some
channels than for others. Completing them is not a new measure or a new cut: the measure stays the same, and the
listing simply reaches every upload it already claimed to count. The comparison is done whatever it shows.

**How.**
1. **Deeper listings.** The six Shorts tabs whose listing may end inside the window are read again, deeper:
   - Behindwoods O2, Chanakyaa, Galatta Voice, IBC Tamil, Minnambalam and Vikatan TV;
   - by `tools/listing.py` (yt-dlp's flat listing, newest first), as before;
   - into `listings-deep/<key>.shorts.jsonl`, the first listings kept as they are.
   - Each is read deep enough to pass 17 Sep 2026 by the dates read.
2. **The window's edges** are found in each deep listing by reading Shorts' own pages for their exact publish time
   (`tools/page.py`), halving between a position known to be inside and one known to be outside, for both edges:
   - the window's end: the first upload dated 16 Nov 2026 23:59:59 IST or earlier;
   - the window's start: the last upload dated 17 Sep 2026 00:00 IST or later.
   - Where the halving stops short of a single position, the sum stays a range over the positions left undecided.
3. **The tab's window sum** is the listed views of the positions between the two edges.
   - It is written to `window-views-deep.json`, with the deep listing's time.
   - The deep listings are read about 16 hours after the first ones, so their views can be a little higher; the
     comparison says so.
   - The long-video tabs and the other Shorts tabs keep their first sums, which are bounded.
4. **The comparison** is made again from these sums whatever it shows, and said as it comes out:
   - who is first;
   - whether Chanakyaa is in the top three;
   - where Vikatan TV stands.

   If a place cannot be told apart within the ranges, it is said as not sure.

### DEVIATION 16, its result (17 Nov 2026, 15:42 IST by the clock; read after the halving, nothing in the rule changed)
- The six deeper Shorts listings (listings-deep/, read 17 Nov 2026 09:56-10:04 UTC): Galatta Voice 4,700 items (a
  first read of 4,000 still ended inside the window, its 4,000th Short from 21 Sep 2026, so it was read again to 4,700;
  the 4,000 read is kept as galatta.shorts.4000.jsonl), IBC Tamil 4,000, Chanakyaa 1,400, Vikatan TV, Behindwoods O2
  and Minnambalam 1,300 each. Every one now ends before 17 Sep 2026; no edge is unbounded (window-views-deep.json).
- The window's Shorts: Galatta Voice 4,255-4,263 Shorts, 81.13-81.18M listed views; IBC Tamil 3,435-3,441, 26.20-26.29M;
  Chanakyaa 790-797, 29.12-29.13M; Vikatan TV 527-626, 7.91-8.75M; Behindwoods O2 747-755, 20.39-20.50M; Minnambalam
  489-496, 19.10-19.16M. (Halving stopped at 6 positions or fewer, as written; Vikatan's start edge stayed wide where
  one page could not be read.)
- Measure 2 (every upload of the two months, long and Shorts, listed views), all 21: Galatta Voice 95.46-95.51M first;
  Saattai 74.11-79.10M second; Chanakyaa 65.09-67.39M and Savukku 59.55-66.23M third and fourth, not sure which (the
  ranges overlap); Behindwoods O2 46.2-46.6M; IBC Tamil 44.5-44.6M; Vikatan TV 25.1-27.4M, 9th to 11th (its range overlaps Neerthirai's and RETROLUXE's).
- The later read: the 600 to 800 Shorts read in both listings gained 0.7% (Chanakyaa) to 1.6% (Galatta Voice) to 6.9%
  (IBC Tamil) in the ~11 h between the reads; Galatta Voice's lead over Saattai's most (16.4M, 21%) does not depend on it.
- What it changes: an earlier reading said Saattai was first by the two months' videos' views; by measure 2 as written it
  is second, after Galatta Voice. The video says the corrected order, measure 2 as written and
  measure 1 unchanged; no cut was chosen after this result (long videos alone, where Saattai is first at 49.23M, is
  not used for the claim).

### DEVIATION 16, corrections after a fresh review (17 Nov 2026, 16:27 IST by the clock; nothing in the rule or the result's order changes)
- The later read: the Shorts read in both listings gained 0.4% (Behindwoods O2) to 6.9% (IBC Tamil), not "0.7% to
  6.9%"; the reads were about 11 hours apart (the "about 16 hours" under How was the estimate before the read).
- The probes were not paced as `tools/window_edges.py` says: each was a fresh `tools/page.py` process, so its pace
  never applied between probes (76 of the 79 gaps under 3 seconds). No page was refused. A probe that could not be read
  leaves no file, so the unread page at Vikatan TV's start edge cannot be shown from the files.
- Published without titles: `listings-deep/` without its title column and `pages-deep/` with dates and counts only, as
  DEVIATION 10 requires of raw listings.
- The second measure counts the Videos and Shorts tabs; live streams are not counted (DEVIATION 1). The video's screen
  says so beside the measure.
- Found in the deeper listings: for Galatta Voice, IBC Tamil and Behindwoods O2 the first Shorts listing, within which
  the kept Shorts were ranked, did not reach the window's start (it reached back to about 5 Nov, about 5 Nov, and
  somewhere between 22 Sep and 9 Oct 2026), and, matched by id and in the deeper read, 115, 64 and 8 Shorts of the
  window that it did not hold have more listed views than the least-viewed kept Short; their titles were not read
  (counts corrected at 16:43 IST: a first count by positions said 116, 75 and 8). Said in RECORD.md, section 8; the kept set and the marks are not changed.
