Ordinary spoken scouting language, stored with a block and a date, contains pest-pressure signals.
This is a look at what's already sitting in the voice notes crews log through Numanac, our voice-powered farm management system.
The setup, out in the field
A two-person crew walks the blocks of a six-orchard operation through the pre-harvest window, describing what they see: no mites, light mites, moderate, high, sometimes with a population read. Numanac attaches an orchard, a block, and a date automatically.
Over six weeks, the crew produced 462 notes. Nobody logged a count or was asked to. The question: is there a pressure curve in there anyway?
Analyzing the language used
Using Alma, Numanac's agricultural intelligence, we scan each note in the record for the pest term.
Rust moderate, more on the north side. Mites light on the south edge, majority still alive, no webbing inside the block. Tree rows clean, a few grass patches at the replants.
Only the bold clause gets "scored": severity light (1), trend building (majority still alive reads as building). Severity is ordinal: none (0), light (1), moderate (2), high or flare-up (3). Of 328 notes mentioning a pest or disease, 273 held a scorable mite clause.
Scoring pest-pressure severity
"Majority dead" and "majority still alive" matter more than "moderate" to someone deciding whether to spray, so we track trend (declining, building, mixed) beside severity.
On Orchard A, severity fell over three weeks — 1.6, 1.1, 0.9 — signaling a population crash. But of the nine notes stating a trend in that window, four said building, three declining, two mixed: the score fell while the scouts' own language leaned toward building.
The transcription trap
Several notes read "2 spotted mites on the edge" — not a count of two, almost certainly the twospotted spider mite misheard as a numeral. We excluded those rather than treat them as either a count or a species ID. "Preparatory mites" were predatory mites; "narcotics symptoms" were necrotic ones. Numanac resolves this domain vocabulary in a layer between the speech engine and the record, before extraction. Otherwise the numbers are wrong in ways no chart reveals.
Adding heat
Pest pressure alone is a list; against heat accumulation, it's a curve. For each orchard we pulled daily temperatures from a public archive (Open-Meteo) at the block cluster's centroid from January 1, computed degree-days by the single-sine method, and summed them.
The thresholds (50°F base, fixed upper cutoff) are platform defaults, not a species-calibrated model. The notes never identify the species with confidence, and published models vary by species anyway. The totals are a consistent, real measure of relative heat. They aren't a precise model of this population.
What the curves showed
Orchards E and F, observed earlier in heat accumulation (roughly 2,600–3,600 degree-days), showed severity rising. Orchard F climbed 0.5 to 0.7 to 1.4 before easing to 1.0 on a six-note week. Orchards A and B, observed later (3,500–4,000), showed severity flat to falling. Orchard D, observed early on just seventeen notes, also carried the highest rate of water-stress language, the condition under which mites usually build. "Early" and "rising" aren't the same claim.
Orchards A and B share a weather grid cell, hence an identical degree-day curve. Orchard C, all disease notes, scores zero for two weeks straight: a zero is still a data point.
The caveats build the takeaway
A spray, predators building, an irrigation change, or ordinary cycling would each fit the same picture, and this data can't distinguish among them. Orchards aren't a controlled comparison — each has its own varieties, history, and spray program. Several weeks rest on a handful of notes, there's no biofix, and each orchard was observed for only three or four weeks. Longer-running studies scale in operational intelligence.
Figures in this piece are illustrative. They follow the shape of a real deployment but do not reveal a specific deployment's data.