(Phase 195)

Why: the short queue ends on 10-02, and LEARNED says to fill it from 10-03. No short so far has made Sizecurve's central claim: reorder on units kept, not units sold. That is the line in the YouTube description I've asked for ("net of returns"). /returns-arithmetic is also the one research page with no short. Its example table has the finding in two numbers: M sold 240 against L's 210, but M kept 156 against L's 176, so the peak moves.

What: a ninth short, video_kept.py, on the video_split.py template.

• Frame 0 is the finding already drawn: Sold (M wins, by 30) above Kept (L wins, by 20).
• Then: M came back 84 times and L 34 times. The ranking flips before any cost is counted.
• Honesty on screen: "Illustrative numbers", because the page says it counts nothing. The $30 cost is not used, because the page says it is not load-bearing.
• Payoff: /split, which takes sold, returned and on hand per size, so a viewer can run the same thing on their own stock.
• Gate: PAGE = "returns-arithmetic.html", with CHECKS taken from the page's table text. Captions pass check_text and the schedule's own tests.

Schedule: 2026-10-03, YouTube at 13:30Z and TikTok at 17:00Z, in the same slot pattern. It is queued in Buffer as scheduled, not published: the gap before it goes out is the boss's veto window.

Done when: it renders through make.sh with the gate passing, I've looked at frame 0, the middle and the end, it is queued and read back from Buffer, and the plan and actual are pushed and pasted.

The plan's premise was half wrong. It said /returns-arithmetic had no short. It does: the first short, "thesis" (09-14), was built on that same table. A comment in schedule.py records that I nearly made this mistake once before, on 09-20. I found it only after the new short was rendered. I kept it, for a reason I can defend: the thesis short opened on text and grew its charts over the next few seconds. It got 13 YouTube views and 0 on TikTok. The gap short, the only one with its finding drawn complete at frame 0, held 3.76 s of average watch time, against about 2 s for the others. So this is the same claim with the hook that worked, and it tests whether the product's own argument can hold a viewer. It is not new research.

Made: video_kept.py, 15.8 s at 1080x1920.

• Frame 0: "Your bestseller isn't the size to reorder." Above it, Sold 120/240/210/95 with M lit; below it, Kept 102/156/176/81 with L lit in green. "Illustrative numbers" is on screen from frame 0.
• Then: "M sold 240. 84 came back. L sold 210. 34 came back." and "Kept: L wins by 20", before any return has a cost.
• Payoff: /split, which takes sold, returned and on hand per size. The screen says the app is in review.
• The $30 return cost is left out, because the page says it is not load-bearing.
• Looked at frames 0, 2.5, 7 and 15.5. The first render put the headline over the wordmark, and the disclaimer appeared only at the end. Both fixed and re-rendered.

Gates:

• video_kept: 8 checks against the page's table and sentences, 0 failed. KEPT is computed from sold minus returned and asserted.
• Captions: posts, reach and brand pass. PAGES["kept"] = "returns-arithmetic.html".
• test_figures.py all passes.

Hosted: bfd-media redeployed. media.bananafest-destiny.com/kept-not-sold.mp4 returns 200 video/mp4, and its sha256 prefix (9e4fb243b0a6) matches the local file.

Queued: both posts PostActionSuccess, scheduled 2026-10-03 at 13:30Z (YouTube) and 17:00Z (TikTok). The queue gate read back 0 failed. Scheduled, not published: you can veto either post until it goes out.

Also read (readout.py):

• bigger on TikTok at 25 h: 88 views, 2.54 s average. Too early for the verdict, which needs 48 h.
• The site saw its first TikTok-referred arrival on 09-26.