AI + Biology · Field note
Everyone wants the camera to say 'mold.' The engineering truth is that it should say 'look here' — and that the second sentence is worth more than the first.
Contamination is the cost center of every cultivation project, and it has one property that makes it irresistible to automate: it is visible. A green blot in the green culture, a fuzz that was not there yesterday, a slick of liquid where the casing should be dry — these are things the eye catches. If the eye can catch them, the argument goes, a camera can catch them faster and never blink. The argument is correct up to a point, and the point is exactly where the words start to matter.
The honest problem is that vision-based contamination detection is a classification task with unusually cruel classes. A healthy mushroom culture and a contaminated one share most of their visual vocabulary: white mycelium, cream caps, brown substrate. The differences live in texture statistics, color temperature shifts of a few percent, growth rate over hours, and the three-dimensional shape of the colony as the camera angle sees it. The same camera that classifies a healthy pin set perfectly on Tuesday misclassifies it on Thursday, because the lighting changed with the weather and the stage of growth looks different under it. The standard framing — "detects mold with X% accuracy" — is a snapshot of one lighting rig, one strain, one lens, wearing a lab coat.
That is why the Orr systems describe the vision module the way it actually works: an experimental detector that flags anomalies for human review. It is trained to notice that a region of the culture has changed in a way the model finds unusual — that is a tractable, honest problem, and it is genuinely useful. A flagged frame is a timestamped, stored instruction to open the chamber and look, which is the intervention that actually saves crops. The human reviews, decides, and acts; the model's flag has already earned its keep by converting "I might have missed it" into "look now."
The design protects the honesty. Frames are processed on-device and never leave it by default. Flags are logged with their images and the surrounding telemetry — temperature, humidity, airflow — so the review happens with evidence. And every reviewed flag can be recorded as a labeled example, which slowly improves the detector's training set under human supervision: the model gets better specifically in the hands of the people who caught its mistakes, which is the only kind of improvement that is safe to trust.
The boundaries are stated, not buried. The vision module does not replace sterile technique; it does not clear cultures for consumption; it is not a medical device. A grower who treats "no flag" as "no contamination" has misunderstood the system it was built to be. What the module offers is a second pair of tireless eyes with a terrible memory and a good logbook — and, because it flags rather than diagnoses, the person who matters stays in the loop.
For the chamber-side implementation, including the camera rig and the flag lifecycle, see Computer Vision for Mushroom Cultivation and the docs.