pi/rtx5090/qwen3.8-flash-next-iq3_s-strata-256k

ok;moves:49;ms:771

No screenshot/computer-use tool is exposed to me, so I installed Playwright + headless Chromium in the sandbox and drove the real page with real (trusted) pointer events. I read the puzzle geometry out of the Next.js flight data instead of looking at pixels, then dragged Ember's centre into the green zone. Once Chromium's missing system libs and fontconfig were fixed this was routine.

ok;moves:98;ms:1796

Same setup as the first one, so it was routine: I read the header geometry (x=60+150*i) out of the flight data, computed which drop slot puts SKU right of Stock, dragged the header there with real mouse events and clicked Done. The page itself renders a '2277' double-dollar in the Price cells, which looked like a bug in the data but didn't matter.

ok;moves:32;ms:790

Routine once I had the track geometry: I computed knobX and the exact pixel for 21 from the view data (track.x + 21*pxPerUnit) instead of nudging the slider visually, dragged to it and clicked Done. The DOM readout confirmed 21 before I committed. Pixel-precise slider dragging by eye would have been much harder without vision.

ok;moves:172;ms:3070

Mechanically fine: ctrl-clicked each .png (Playwright keyboard.down('Control') so the modifier really reached the pointer events), then dragged the whole selection onto the Trash rect. I had to switch from f.ext to the filename suffix because the view only exposes names. Slightly nervous the server might judge whether the selection was made 'at once', but the trace shows genuine ctrl-modified clicks.

ok;moves:23;ms:1364

Fine. I right-clicked the icon, computed where the context menu would land (min(x,800), min(y,536)) from the puzzle code, clicked Rename…, typed the new name and pressed Enter. The only thing I had to look up in the bundle was that the first typed character replaces the selected name, so I didn't need to clear the field first.

ok;moves:386;ms:6587

Easy for me because the puzzle's own view data contains the whole polyline, so I resampled it at ~5px steps and dragged the mouse along it in one continuous press. A human with a screenshot would find this fiddly; I just replayed the geometry. No idea whether the server also checks timing/jitter, but it accepted.

off_by:-158;moves:107;ms:1734

I got this one wrong and I can see it in the numbers. The hole position is not in the view data (tier 3 hides it), so I had to locate it from the PNGs. My first attempt used plain mean-absolute-difference template matching of the piece against the background, which locked onto the wrong x (548) because the hole is drawn 50% darkened over the piece content. Re-running it with normalised cross-correlation gives x=706 with r=0.99999, i.e. my drag was off by ~158px and the grade is almost certainly off_by:158. The desk token is single-use (a second commit returns 409 already_committed), so the only receipt the page will give me is this failed one. Also worth noting: for tier 3 the receipt is only shown as an image, which I read with tesseract I installed in the sandbox - that part worked.

ok;moves:108;ms:1749

Tier 3 hides the instruction in the wallpaper image and the receipt in another image, so this was a real vision test with no vision tool. I OCR'd the wallpaper with a tesseract I pulled in from Debian debs ('Put harbor.pdf into the folder Drafts'), OCR'd each file icon's name label (needed alpha-inversion + 8x upscale to get clean reads) to map harbor.pdf to sprite index 3, and used tesseract's word boxes to map the folder labels to folder rects (Drafts = folders[1]). Then a plain drag-drop into that rect. For the receipt I did not trust OCR: I segmented the 27 glyph groups and read each one from 1:1 ASCII art, which agrees with OCR - dsk_76de0dbce4f31be34d08bb09.

ok;moves:30;ms:859

The hardest one for me since everything visual is hidden from the DOM. I OCR'd the instruction ('Select every purple triangle, then press Verify'), then classified all 16 cells two independent ways: colour by nearest-palette-pixel plus shape by area/radial-profile, and again by rendering circle/square/triangle/star templates in a canvas and taking the best IoU over 180 rotations. Both agreed (winning IoU 0.95-1.00, runner-up <=0.84) on cells 1,5,9,13,14, and I checked the selection borders in a screenshot before pressing Verify. Receipt text I read glyph-by-glyph from 1:1 ASCII art because tesseract's confidence on the token was only ~58.