Reading the electrome

Three mechanics from the electrome literature, each shipped at the size the evidence actually supports.

🏷 Intervention signatures — labels nobody had to be asked for

The published stress classifiers reach high accuracy on labelled data, inside one experiment, on plants somebody deliberately stressed in a known way. That is the part that does not survive contact with a houseplant: nobody at home can say "this window was salinity", and a model pre-trained on someone else's greenhouse would be confidently wrong about your ficus.

But the labels are not missing. Every watering and feeding is already in the care log with a timestamp. So instead of classifying stress the plant might be under, this learns the signature of things known to have been done to it:

await plant.learnInterventions()
// 'Characterised 4 of 4 logged interventions from the electrode buffer.'

plant.interventions.identify( samples, rate )
// { match: 'water', confidence: 0.82 }

An intervention whose own occurrences look nothing alike has no signature, and is reported as unreliable rather than averaged into a shape none of them has. Two equally close candidates are reported as a tie, not resolved by picking one.

📊 Regime change

Not "does this window depart from normal" but "is there a point in the record on either side of which the plant behaves like two different systems".

The temptation is to promise lead time — detects drought three days before wilting. Nobody can promise that here; the published lead times come from deliberately stressed plants under instrumentation nobody has at home. This reports that the regime changed and when. Whether it precedes anything in your plant is something only your plant can eventually tell you.

🏘 Collective state

A colony-level signature is not evidence that plants influence each other — they share a window, a radiator and a watering can, and that explains nearly every correlation you will see. Reading coordinated change as communication is the easiest mistake here, and it is unfalsifiable without a sensor for the supposed channel.

The opposite inference is sound: several independent plants shifting at once is strong evidence of a shared environmental event, caught more sensitively than any one of them could manage. One plant changing is a plant. Every plant changing is the room.

🦠 Early infection — the one inference that earns two modalities

A drop in electrical complexity and the first yellowing seen by a camera do not share a failure mode: a bad electrode causes no chlorosis, and a white-balance error lowers no entropy. When both move together they really are two witnesses.

It reports; it does not act. Isolation and treatment are physical and expensive to get wrong, so they stay behind the risk gates — an automatic quarantine on a false positive is worse than a late true one. And if the electrical half is a single drifting electrode, the whole thing is refused.

⏱ Fresh data for anything that acts

Blue light forces stomata open, so the point of gating it on VPD and soil is knowing what the plant faces now. The contraindications already refused absent data; a three-hour-old reading is worse than absent, because it looks like knowledge. Stale readings are now treated as no readings.

  1 min → allows
 45 min → 'The readings gating this treatment are 45 minutes old, and a
           treatment that forces a physiological response has to be decided
           on current conditions. Take a reading first.'

Probes are unaffected: reading the stomata does not force them.