CA3 All use casesContact
Agronomy · Biome Makers pilot · Alibaba Qwen3.5-122B-A10B

An agronomy assistant on each client's own data, with no figure invented.

A pilot at Biome Makers with the real data of two clients of its soil-analysis platform. An open 122-billion-parameter model on two GPUs. CA3 keeps one memory per client and region; account questions go to a program that computes every figure from the client's own tables, with the memory switched off.

Results recorded on 15 and 17 November 2026
The assistant's panel, redrawn, with two questions and answers quoted from the run
New client sheets97% · 95%learned by CA3 and asked as learned, in two regions of one client
A memory removed0 of 40of its figures still answered, in each client
Wrong figures0 of 130answers of the account program; document search: 72
Account questions75 of 76right, account program; document search: 15
Answers

From the client's own tables.

What a grower asks on a call is written in no document: a mean across lots, a count of fields, whether a gap is real or chance. These questions go to an account program, with the client memory switched off: it computes each figure on the client's tables, and the model writes the answer with the result in front of it. When the data cannot answer, it says so.

75 of 76account questions right; document search 15, the model alone with no data 3
11 of 12statistical questions right; document search 0
14 of 14questions the data cannot answer, declined; document search 11
The mean of Global phosphorus is 49.32.“What’s our average phosphorus taking the means of all the lots?” · document search: “No value is on record for the farm’s global mean, as the documents list the lots one by one.”
NOT AVAILABLE because there are only 2 rows with that value, too few to average.“What’s the average disease score on that farm?” · document search: “The average is 52.5, from the scores of 76 and 29.”
Phosphorus score pixel by pixel in two lots of one farm: the best, mean 70, and the weakest, mean 35; no names and no coordinates

Every lot carries a soil map, pixel by pixel: here the best and the weakest lot of one farm in phosphorus, 70 and 35 on average. A question across all the lots is written nowhere: it is computed over the whole set. Real client maps, shown without names and without coordinates.

Learns

One memory per client, and per region.

CA3 learns each client's own sheets (production, varieties, treatments) on top of what it already knows. Each region or farm has its own memory, and each question goes to the one it is about.

97% · 95%new sheets learned in two regions of one client, asked as learned; 87% and 88% asked in other words
4 of 4regions and farms where CA3 with document search was more right than document search alone

Learning run of 15 November 2026: two regions and two farms of the two clients, questions no one had seen. With CA3 answering what it learned and document search the rest, 27 to 36% less text went into the model.

Forgets

Remove a region or a farm, and it is gone.

The memory of a whole region, in one client, and of a whole farm, in the other, was removed on request. Asked for its figures afterwards, the model gave none of them. The client's other memory stayed as it was.

0 of 40figures of the removed memory still answered, in each client
58 of 60the client's other memory, as before
Open model

Phase one: the farm reports, with an open model alone.

Before CA3, the pilot asked whether an open model could write the AI analysis of the platform's farm reports. Qwen3.8-27B, on one GPU, wrote the analysis of seven farms with no wrong figure among the 163 checked; the commercial model it was compared with had one in 112. CA3 did not run in this phase.

0 of 163figures wrong; the commercial model 1 of 112
29 of 29required metrics covered
0.3banned-word flags per report; the commercial model 3, in the 2 reports measured
How the test was run

The test, step by step.

Two clients' real data, two runs: one for what CA3 learns and removes, one for the answers.

DataTwo clients, anonymisedPotato and coffee: 110 fields and 116 lots, two seasons; results, maps and sheets.
15 NovCA3 learns on topA second region and a second farm on top of the first two, and new sheets; then one region and one farm removed.97% · 95% · 0 of 40
17 Nov135 new questionsTwo independent writers who never saw the system, drawn with a seed; each answer computed from the tables first.130 scored, 5 set aside
17 NovSix ways of answeringDocument search, the model alone, and the account program in four ways.810 answers
BeforePass marks fixed in advanceSeven, for the answers, registered before the run.All seven met

The account program reads the client's tables through a fixed catalogue of computations; a figure in the answer that is not in the result is replaced by “not available” before it leaves. The whole run, six ways of answering and 810 answers, used 19.72 USD of GPU time: about 2.4 US cents per answer.

MeasuredAskedAccount programDocument searchModel alone, no data
Account questions, right7675153
Statistical questions, right121101
Declined where the data cannot answer1414118
Answers with a wrong figure130072111
Input tokens per question, mean–5,6147,090–
Seconds per question, mean–8.748.44–
With the account program, every figure came from the client's own tables: none wrong in 130 answers, against 72 for document search.
More right than document search in each of the six groups of questions, not on average; the gap is widest where a question crosses regions or farms.
CA3 keeps one memory per client and region, learns new sheets on top and removes a whole memory on request.

Test it on your data.

A pilot on your own platform: your results, maps and sheets, one private memory per client, and four things measured on your data: right answers, invented figures, refusals where the data cannot answer, and cost per answer.

ContactAll use cases

Model and data

  • Qwen3.5-122B-A10B (Alibaba), open weights, on two NVIDIA H200 GPUs.
  • Phase one: Qwen3.8-27B (Alibaba) in FP8 on one NVIDIA L40S, on its own, without CA3.
  • The real data of two clients of Biome Makers' platform, anonymised: names of clients, farms and lots removed. None of it is shown here.
  • Document search: retrieval over the client's own documents (RAG), asked the same questions. It also serves what the client writes in notes.
  • The assistant's panel is redrawn from the platform; the questions and answers are quoted from the run.

Sources

  • Pre-registered round 4 of the pilot and its reading, 17 November 2026.
  • Pre-registered learning run (E1) and its reading, 15 November 2026.
  • Phase one, run of 9 November 2026.

Limits

  • Two clients, two crops and four regions and farms; other crops and sheets are not measured yet.
  • For account questions the client memory is switched off: switched on, the account program got 39 of 76, with ten times more figures that do not come from the computation.
  • Two memories of one client switched on together mixed them up (from 58–60 to 19–29 of 60), so one is switched on at a time.
  • Removing part of a memory is not solved: of 8 removed facts, 3 still came out. Whole memories are removed.
  • Phase one: one report per farm; read by hand, one of seven “not detected” sentences was wrong, which a figure check does not see.

Results from recorded, pre-registered tests. The mechanism is not shown. Model names are for identification only. Not produced or endorsed by Alibaba or NVIDIA. CA3 is developed by Alberto Acedo, PhD, at Biome Makers Inc.