The chatbot was never in the tent

The grow boards have settled on a view of AI this autumn, and it’s earned. Someone posts a photo, a chatbot gave them an answer, and the answer was confident and wrong. The replies say to stop asking it.

Look at what that chatbot actually had to work with. One photo, taken under whatever light was on. Whatever the grower remembered about temperature, humidity and watering, typed from memory. Nothing from yesterday, nothing from last week. It has never been in the tent, so it fills the gaps with whatever sounds most likely. That’s what it’s built to do.

So the problem isn’t the model. It’s the inputs. Croplock is built the other way round: the box lives in the tent, and the question gets asked of the record, not of somebody’s memory.

It measures the leaf, all day

The pods log air temperature and humidity every minute, and an infrared sensor reads the surface of a leaf next to them. That gap between leaf and air matters more than any one reading. A well-watered leaf that’s transpiring runs a little cooler than the air around it. A plant that has stopped transpiring runs warmer, and under some conditions well above the air. Nobody can tell you that number from a photo.

Yesterday, day 98 of the run in our tent, the leaf sat a median of 0.94 °C below the air while the lights were on, never further than 1.25 °C below and never above it, across 918 paired readings. That’s a plant drinking. It’s one plant on one day; what makes it useful is that the same comparison has been running in our tent since July, and the record shows the few times it stopped.

It writes everything down

Every reading, every hourly camera frame, every time the fan or the humidifier switched in our tent, every phase change and every watering a grower enters lands in one record on the box, dated. When something looks wrong, the first question isn’t “what do you remember?” It’s “what does the record say happened?”

That’s not a nice extra. In July a seedling in our tent went yellow at the tip. Its history showed nine straight days at 76 to 82% humidity just before, the stretch when that tissue was built. A chatbot would have been shown the yellow leaf. The record showed the nine days.

Small models, each with one job

The camera doesn’t hand a photo to a general model and ask what’s wrong. It runs small models on the box’s own processor, each trained for one narrow question. Is there a plant in frame. Are pistils present. Are pollen sacs present. Is this worth a person’s eyes. Each one is trained on the job it does, and each can decline to answer when the picture doesn’t support one.

On day 84 the pistil check read the hourly close-up of the top of the plant, said pistils present seven hours running, and the hub wrote FEMALE into the grow’s record on its own. It also checks for pollen sacs every lit hour; yesterday that was 1,215 reads, every one absent.

Advice with a source, and a loop that acts on numbers

When the box gives advice, it comes from a written body of cannabis facts with the source named, and when that source has nothing to say, it says so instead of improvising.

In our tent, the box also switches the fan and the humidifier, on measured numbers against the targets for the current phase. A chat has no part in that. The language model only talks; it has no way to switch anything.

IG-047What each one has to work with

The same question about a plant, asked of a chatbot and of the box in the tent.

A chatbotThe box in our tentThe plantOne photo, under whatever light was onAn hourly frame, kept in the recordThe climateWhat the grower remembersAir temperature and humidity every minute, and the leaf beside themLast weekNothingEvery reading, frame, switch and phase change, dated, in one recordThe answerWhatever sounds most likelyA cited fact, or a plain 'nothing to say'When it is wrongThe chat closesThe miss is dated, next to the readings that explain it

It is still wrong sometimes

We’re not going to claim this box can’t be wrong. It has been, and it’s on the record.

The camera called the plant female ten days after a person did. As that post put it: “A person saw it first; the machine wrote it down second, with seven hourly reads behind the row.” A first pistil is under three pixels wide in this camera’s view, so the machine will always trail a close human look.

And the yellow seedling? The tent was inside the humidity range the device’s own pack set for a seedling the whole time. The device did exactly what it was told, and the plant got sick anyway. It took a person reading the record to see it, and the pack’s seedling range was changed because of it: after the first week it now eases toward the drier range a vegetative plant gets.

That’s the real difference. When a chatbot is wrong, the chat closes and the mistake goes with it. When this box is wrong, the miss is dated, sitting next to the readings that explain it, and someone can find it and fix it.

The chatbot was never in the tent. The box is in it every minute.

Sources

• Nelson, J.A. & Bugbee, B., “Analysis of Environmental Effects on Leaf Temperature under Sunlight, High Pressure Sodium and Light Emitting Diodes”, PLOS ONE 10(10): e0138930, 2015. Why a transpiring leaf runs cooler than air, and a stressed one warmer.
• Day 98 leaf and air readings, pollen-sac reads: our own hub’s record, read 2026-10-16. First-party.
• The yellow seedling and the nine humid days: In band, and sick anyway.

Also read: The growers are right about the chatbots · In band, and sick anyway · The camera called it female, ten days late · The whole run, as the device recorded it