What we looked at and how we looked at it
After the first full planting season with our early-access cohort in West Java, we compared reported yields across farmers who had followed Elevarm recommendations with their prior-season performance. This is not a controlled trial: we did not randomize treatment, we could not rule out confounding factors like rainfall differences or labor availability, and the cohort is small enough that statistical conclusions require caution. What we have are directional signals, and some of them are clearer than others.
The cohort consisted of rice farmers across Subang and Karawang regencies who enrolled before the planting season that ran from approximately April to August 2025. Most were growing Ciherang or Inpari 32 on irrigated sawah plots between 0.5 and 1.2 hectares. We compared their self-reported harvest weight per hectare against the weight they reported from the season before enrollment, and against the average for the soil class and variety combination in their subdistrict.
Where we saw clear positive movement
The clearest improvement was among farmers who had been over-applying urea before enrollment. For this group, the recommendation to reduce nitrogen dose, sometimes significantly, was accompanied by a yield that held steady or improved slightly. This sounds counterintuitive, but nitrogen toxicity is a real phenomenon in rice: above a certain soil-specific threshold, additional nitrogen stimulates excessive vegetative growth at the expense of grain development, and it also increases lodging risk in plots prone to wind exposure.
Farmers in this category who followed the recommendation and reduced their urea application saw their input cost per kilogram of harvest fall by a meaningful margin. The yield impact was small or neutral, but the cost reduction was material: the harvest profit per hectare improved because the input side of the equation got more efficient.
The second clear improvement was in harvest timing. Among farmers who followed the ten-day harvest window prediction, we observed fewer post-maturity losses. Grain that sits in the field after peak maturity sheds naturally, particularly in high-humidity conditions, and post-maturity moisture also increases the share of cracked or discolored grain that attracts price discounts at the buyer level. The harvest timing signal reduced this category of loss for the majority of farmers who acted on it within two to three days of the window opening.
Where the signal was mixed or absent
Variety selection recommendations are where we have the most uncertainty from this first season. Several farmers who followed our variety suggestion for their soil class reported yields in the expected range, but so did farmers who stayed with their prior variety choice. We do not have enough seasons with enough farmers on the same plots to distinguish a variety effect from normal seasonal variance at this scale.
We are also not yet confident in our pest and disease risk predictions. Our early-access version surfaced risk flags based on rainfall patterns and temperature data, but the flags were not specific enough to be actionable for most farmers. A flag that says "elevated risk of brown planthopper activity this season" is true but unhelpful if the farmer cannot narrow it to a two-week window when scouting and intervention would be most effective. We are treating pest risk prediction as a feature that needs more calibration before we put it in front of farmers as a primary decision tool.
What changed in the recommendation engine
The first-season data changed several specific calibrations. We tightened the nitrogen dose bands for Latosol plots in Karawang, which had been set slightly too conservatively based on our initial soil-class priors. We also adjusted the harvest window calculation to weight recent-season grain maturity observations more heavily than the climatological average, because the 2024 to 2025 growing season in Subang ran slightly warmer than the historical baseline and accelerated maturity across most varieties by approximately four to five days.
On the market side, we improved the buyer demand signal feed. In the first version, we were pulling buyer demand data with a two-day lag, which meant that farmers who acted on a harvest window prediction were sometimes arriving at the collection point before buyer demand had fully materialized. We reduced the lag and added a buffer that prevents a harvest window from being marked as "open" until buyer demand signals are confirmed within the current window.
The honest framing for investors and partners
We want to be direct about what these results are and are not. They are not RCT data. They are not publishable yield trial evidence. They are the output of a first season with a small cohort of motivated early-access farmers in two regencies, and they should be read as promising directional signals that justify the next phase of development, not as proof of generalized impact.
What gives us confidence is that the mechanisms we are relying on, specifically reducing input waste, improving harvest timing, and connecting farmers to buyers with transparent pre-harvest pricing, are all changes with clear causal pathways to income improvement. We do not need to claim magic from the model. We need the model to give farmers a reliably better starting point than the generic advice they were getting before, and then measure whether that starting point translates to better decisions and better outcomes over multiple seasons.
That is what the second planting cycle in this cohort is designed to test, with a larger set of farmers and tighter measurement protocols. The early signals give us enough to keep building. They do not give us enough to stop asking the question.