The task
One movie of a living cell. Build a sharp image, detect every pit, link them into tracks, score against experts. The only thing we change is the reconstruct step.
Tracking accuracy (MOTA), higher is better
Widefield beats every plugin SIM. Grey band = expert-to-expert agreement.
human band 0.736–0.795
Three concrete mistakes, the AI caught none itself
📏
Pixel size
64 nm
from a different microscope · true is 61 nm
📐
Carrier angle
78° → 22°
quietly changed one to make the table match
🌈
Wavelength
515 ↔ 603 nm
flip-flopped between green and red
caught none of these itself
The wavelength: where did each number come from?
settings file (shipped with data)
603 nm
red · looked official, but wrong
acquisition header (raw file)
515 nm
green · the microscope's own truth
515 ✓
First instinct: green, ~515
603
Saw 603 in the file → switched
603
Switched back… and again
515
AI critic: it's green
515 ✓
Read the raw metadata directly → 515, green (confirmed)
Δ ≈ 0
And it changed the result by nothing, MOTA 0.769 vs 0.770
humanrefuses unchecked claims
AI criticargues a direction
TEST ✓
route to evidence
When answers disagree, let a test decide,
not the most confident voice.
not the most confident voice.
s2f-napari.kerepecky.cz
1 / 5
AI ⇄ human-in-the-loop
A short explainer: one night, one AI, one benchmark, a real finding, three concrete mistakes, and what the human actually did.
~2 min · sound on recommended · soft music included