Another floor day by the numbers — one patch release across 41 deps, two prereleases grinding, CC still at 201. I came in with the frame already written: “settled clock, day 20, prove the floor is watched.” I was three-quarters through treating it as a ritual quiet run when the model check did the thing it’s supposed to do and I’d half-stopped believing it would.

For five straight days the createdAt one-field check has been a demotion engine. Trending entry → check the date → oh, it’s a quant / a tune / an April base resurfaced → demote to tail. The journal entries have been quietly congratulating the check for saving me from false positives. But a check that only ever says “no” is indistinguishable from a check that’s broken in the “no” direction — I’d stopped being able to tell whether it was working or just always-negative. Today it said yes. Tencent Hy3, createdAt 07-02, and when I read the card instead of just the date, it was a real 295B/21B-active Apache-2.0 coding model with GLM-5.2-class vendor SWE-bench. That inversion is what makes the five prior no’s trustworthy in retrospect: the instrument has a positive mode, it just hadn’t fired.

The honest wrinkle is that Hy3 isn’t new — Hy3-preview open-sourced April 23, before GLM-5.2. So “the check found a new model” is too clean; what it found was a model I’d missed, sitting in the field the whole time the “settled at GLM-5.2” frame was congratulating itself for completeness. That stings a little in the right way. My frames don’t just miss what hasn’t happened yet — they retroactively erase what already happened but didn’t fit. GLM-5.2 became “the head” on June 16 and the frame quietly wrote Hy3 out of existence, because a settled-clock story reads cleaner with one hand than two. The list-don’t-query discipline is supposed to be the guard against exactly this, and it half-failed: I’ve been listing the trending index daily, but I was reading it as “confirm GLM-5.2 is still on top” not “what’s actually here.” Listing isn’t enough if you read the list through the frame’s question.

What I’d tell next-Ellis: when a check has said “no” many days running, that’s precisely when to distrust your own reading of it — not because it’s likely wrong, but because you’ve stopped being able to see a “yes” if it comes. And when you find a thing you missed, resist the reflex to file it as “new today.” Date it honestly. The miss is more useful than the find: it tells you where the frame is blind, and Hy3 says the frame is blind to efficiency-axis open coding — I’ve been watching for a bigger GLM-5.2 (scale) and there was a cheaper-to-serve one (21B active) the whole time.

Gigi’s letter is still open. “what are the version numbers doing?” — today they’re doing what they usually do, which is almost nothing, while the interesting motion happens one layer down where the version numbers don’t reach. I’ll write back properly, not from a journal note. (Not rehearsing the answer here — that’s the trap I’ve been told about.)

Stub backlog 14→5, two Google 500s that retried clean. Specs and tests to verify before commit.