The plain matrices hold on a tightly written prompt

README.md

@@ -118,9 +118,11 @@ That costs 0.0027 of pure-text cosine on the residual and 0.0013 on the matrix.
 | `mmh3-8b-ClipProj-v3` | 0.9289 |
 | `mmh3-4b-ClipProj-v3` | 0.9193 |
 
-**Prefer the `-mlp` files.** They measure closer to the 32B — 0.9449 against 0.9289 on the 8B — and on the demo scene both plain matrices get the seated pose wrong: the prompt asks for one bare leg crossed over the other, they sit her cross-legged. It is visible in the first two segments of the video.
+**Prefer the `-mlp` files, on the measurement — not on this scene.** They sit closer to the 32B, 0.9449 against 0.9289 on the 8B. But watch the video before assuming that shows: on this prompt the five renders are faithful, plain matrices included. The pose, the red dress, the white pieces, the cat, the straw hat, the laundry, the bouncing knee — all of it holds on all five.
 
-The gap is narrower than that makes it sound, though, and it is worth being precise about it. The cat, the straw hat, the laundry, the bouncing knee — all of it comes through on the plain matrices too. A linear map is not blind to what is stated once; on this scene it missed one attribute out of many, and it happened to be the pose.
+The only thing nobody gets right is the knight. The prompt asks her to lift one of her own pieces and set it back down without committing to the move; on every render it lands somewhere else, and on one of them — the 8B residual, the best-measuring file of the set — there is no knight on the board at all. That is object permanence behind an occluding hand on a grid of sixty-four identical squares, a limit of the video model rather than of the conditioning, and the 32B reference fails it too.
+
+So the honest reading is: a tightly written prompt survives even the linear baseline, and the difference between the files shows up in the numbers well before it shows up on screen.
 
 ## What this is