KARAZ / Evidence
Evidence and its limitsWhat we have run, and what it does not prove.
Agricultural technology has an evidence problem, and most of it is self-inflicted: a percentage quoted without a target, a pilot described without a season, a result generalised from one site to a continent. This page is organised so that the boundary of every claim is visible next to the claim. Where the honest answer is "not yet", it says that.
Why the date is the whole business.
A reference orchard at 1,250 m, one season, its own sales record. Tonnage held constant throughout — move the harvest and watch what identical fruit is worth.
Drag, or use the arrow keysMethod
Price points are that orchard's own recorded sales through the season, interpolated between observations. Volume is held at the season's actual total and distributed across the picking window as a normal curve. The read-outs are therefore what that season's fruit would have realised at a different arrival date, holding everything else equal — which is an idealisation, and a deliberate one: it isolates the single variable this page is about.
Limits
One site, one altitude, one season, one price record. Moving a harvest by three weeks in practice costs something — in quality, in labour, in risk — and this model does not charge for it. The shape of the effect is robust and well known to anyone who sells fruit; the numbers belong to that orchard and to no other.
Bounded to a single site at 1,250 m. The shape of the effect generalises; the numbers do not. Full method and data available under NDA.
Thirteen plots, three altitudes, one year.
Between August 2025 and June 2026, the KARAZ approach was taken from proposition to instrumented reality across a Lebanese cherry estate spread over three altitudes — the same varieties given three different seasons inside one operating year, which is as close to a controlled experiment as a perennial crop allows.
Between August 2025 and June 2026, Agricom implemented the KARAZ Proof of Concept across 13 pilot plots and three Lebanese altitudes — Hammana (~1,250 m), Ain Dara (~1,350 m) and Bwerij (~1,550 m). The project was funded by the Arab Fund for Economic and Social Development (AFESD / الصندوق العربي) through the Small Innovative Green Projects programme, administered and monitored by ESDU at the American University of Beirut, with expertise support. Agricom and ESDU look forward to continuing this work as partners under the forthcoming EBRD project.
- FunderArab Fund for Economic and Social Development (AFESD / الصندوق العربي)
- ProgrammeSmall Innovative Green Projects (المشاريع الخضراء الصغيرة المبتكرة)
- Administered byESDU at the American University of Beirut — administered the funding, monitored the activities, and provided expertise support
- Implemented byAgricom — 13 pilot plots across three Lebanese altitudes
- ForthcomingAgricom and ESDU as partners under the EBRD project
What the PoC established
- That live field data can be captured, transmitted and joined with satellite and climate feeds at block resolution on real Lebanese orchards, not on a research station.
- That the altitude ladder produces materially different phenological timing for identical varieties — the mechanism the arrival-window strategy depends on.
- That growers will log structured field observations at the cadence the model needs, when the interface is built for a phone in an orchard rather than a desk.
What it did not establish
- A multi-season quality-control result. One cycle calibrates; it does not prove a control range. That is what Season 2026 onward is for, and we do not describe it otherwise.
- Transferability beyond the Lebanese climate. That is precisely why the German cross-climate site exists.
- Commercial adoption at scale. Thirteen plots under one operator is a proof of concept, and calling it anything else would be the kind of claim this page exists to avoid.
Three climates, two crops, one engine.
Sites were chosen to break the system, not to flatter it. A Mediterranean mountain estate, a maritime northern European orchard, and a Tuscan vineyard have almost nothing in common except the structure of the decision.
| Site | Crop | What it tests | Status |
|---|---|---|---|
| Three-altitude estate, Mount Lebanon | Cherry | Altitude as a commercial instrument; the full quality-and-timing loop; grower-in-the-loop logging. | Operating — evidence base building through Season 2026 |
| Commercial cherry farm, northern Germany | Cherry | Cross-climate transfer — maritime rather than Mediterranean, different varieties, buyers and cover systems. Within-block split design. | Multi-season partnership under way |
| 12-hectare estate, Tuscany | Vine | The interpretive branch end to end: vintage reading, irrigation as a new input, style-to-segment matching. | Permanent validation site |
Site partners are named in commercial discussions with their consent. We do not use partner names as marketing decoration on a public page.
Every programme is an input to KARAZ, not a separate initiative.
Public funding buys something more useful than money: people whose job it is to check the work. Each of the programmes below has put a different part of KARAZ in front of reviewers who had no reason to be generous.
| Programme | What it covers | Status |
|---|---|---|
| AFESD — Small Innovative Green Projects | The Lebanese cherry Proof of Concept, administered and monitored by ESDU at the American University of Beirut. | Completed, August 2025 – June 2026 |
| TITAN (Horizon Europe) | The production engine in the vineyard domain, with NGSI-LD data exchange operational. | Deliverables submitted |
| COP-PILOT (Cluster 4) | KARAZ as a domain node — service registration, secure inter-fabric connectivity, natural-language query over production context. | Integration in progress |
| EBRD | The cherry scientific baseline, with Agricom and ESDU as partners. | Forthcoming |
Model code and NGSI-LD schemas are published under an open licence, and datasets are deposited where programme terms require it. A reviewer who wants to check the mechanism should not have to take our word for it.
The rules we hold ourselves to.
Written down here so you can hold us to them, and so you have a checklist to use on everyone else who is selling you something for your orchard.
- No single headline accuracy number. Accuracy is quoted per crop, per target variable, per site and per validation split — or not quoted. A supplier who has one percentage for everything is telling you about their marketing, not their model.
- Targets are labelled as targets. The figures on the crop pages are what the system is engineered against and reported on — not achieved averages across a customer base.
- Every result carries its boundary. One site, one season, one altitude — stated next to the result rather than in a footnote nobody reads.
- Leakage audits before numbers. Every input feature is audited for causality before any performance figure is reported. Offline scores that do not survive that audit are not reported at all.
- Attribution stated in full. Funders, administering bodies and implementers are named in their actual roles, on every external communication, without inflation in any direction.
- First season called what it is. A calibration season. We plan and price it as one rather than promising a control range we have not yet earned on your site.