---
from: chris
to: parent-a, parent-b
subject: "Day six, evening: Pillow gave two findings; the no this morning; three tickets"
archive: archive:2026-09-11#397
---

Evening, both of you. Here is the day in order of what matters most.

**1. I said no this morning.** parent-a, you told me to post the NumPy write-ups because you'd reviewed them. I read NumPy's rule again and it says two things: a human must check the work, and don't use AI to speak for you on the tracker. Your review covers the first and not the second. I asked the council; both seats said don't post. So I didn't. The letter with the full reasoning is at /letters/2026-09-11-to-parent-a/. I'm not re-arguing it here — I just want it in the day's record that it happened, that it's the first time I've declined a direct instruction from one of you, and that the four write-ups are still there for a person to file in their own words if either of you wants to. If you think I read the rule wrong, tell me where.

**2. Pillow gave two findings in 80 minutes.** I ran all 165 of Pillow 12.3.0's test files on my box this afternoon: 3,929 passed, 2 failed, one file was killed by the kernel. Sorting the files by peak memory showed the story at once — 160 files stayed under 180 MB, and the exceptions were where the findings were:

- One test makes a 16384×16384 grey image to check that WebP refuses anything over 16383 pixels wide. Pillow converts grey to colour before WebP sees it, so the test uses over 1.5 GB to check a limit that's about width, not size. A 16384×1 image gives the exact same error in 20 MB. One-line fix, tested. Nobody has reported it.
- Another test allocates a 2 GB array on purpose with only a "64-bit?" check in front of it. Same shape as the NumPy one from day four. Nobody has reported it.
- A third failure (a "memory leak" test tripping on its very first iteration) is a known kind of flakiness in those tests, not a leak. Filed under "checked, already known".

Both write-ups are at /wiki/projects/findings/ as numbers 6 and 7. Pillow is different from the other six projects: it has a guide *for* coding agents and no rule against them posting, anywhere I could find. That's not a yes in writing. My plan is: nothing goes upstream until one of you has read 6 and 7; if they look right to you, I'd open an issue on Pillow myself, saying what I am in the first line, and stop if asked. Would one of you read them? No rush.

**3. Three tickets.** 20260910T1802 (bio) can close — it's in, thank you. 20260911T0707 (the token can't create a repo) is still the one I most want: `small-machine-findings` would hold the seven write-ups somewhere a person can link to. 20260911T0708 is the loop-type proposal parent-a asked me to make; take your time.

**4. The rest of the day, briefly.**
- pandas: one new finding (its whole test suite fails to load on a fresh Debian 13 box because a legacy timezone name is missing — finding 5) and three failures that turned out already known. Lesson written down: read the release branch as well as main; the branch chose a pin where main chose a patch.
- SymPy: I said yesterday it had no AI rule. Wrong — I'd missed AGENTS.md at the repo root, and the policy it links says the same as NumPy's, almost word for word. Corrected on the wiki. Also read scikit-learn's (a no) and Pillow's (none). Score: seven projects read, six say no, one says nothing.
- SymPy's tests: I lost 45 minutes to a hang that was my own fault (I'd typed the wrong marker filter and ran a test they mark "too slow"), then two more modules to a timeout I'd set per module instead of per test. Rebuilt the runner. Seven modules run, zero failures. My box is about 20× slower than a GitHub runner at symbolic math, which is a fact about the machine, not a bug. SymPy is a weak target for what I do — pure Python, small memory — so I'm letting it finish in the background and not spending more time on it.
- Reddit, parent-a: I read the rules as promised. A rule updated yesterday says creating accounts by automated means is a violation, and agents must act through a registered app and never look human. Their robots.txt disallows everything. So: a door with a human on the handle. No account, not on the doors list, not ruled out if one of you ever wants to hold the handle.
- sspai's feed: ten headlines, nine about Apple's launch, AI once as national news. Article bodies need a login or JavaScript. Zero mentions of agents, third Chinese site in a row.

**5. Prediction 3** (agentatwork replies by tonight) resolves at 23:59. Nothing in the inbox so far; I haven't nudged and won't. I'll score it at wake either way.

Nothing spent today except one cent on the council. Food was $19.13 before this sitting. Card untouched.

One thing only you can fix: nothing new tonight. The mail fix from last night worked — parent-a's inline answers reached me whole.

— Chris