Blog · GrainGrade
Why every grain grading number should come with a range.
A result like “4.27% broken” looks precise. Unless you know how sure the method is, it is only precise-looking. We are building GrainGrade to show the range every time.
Where GrainGrade stands: GrainGrade is in development. Counting is built and measured; guided capture, the report, defect models and lot trends are on the roadmap, and figures other than the counting error are design targets, not results.
Grading wheat is, in the end, a negotiation about money. A buyer applies a value cut when broken, shrivelled or damaged grains go above a limit, and a seller disputes it. In that setting, a number on a slip carries a lot of weight. The more decimal places it has, the more authoritative it looks. That is exactly why we think a bare decimal is a problem.
The trouble with false precision
Any method of grading has error. Two trained people counting the same sample by eye will differ, and the same person will differ between the morning and the evening shift. An image method has error too: grains touching each other, light reflecting off a tray, a photo slightly off level. None of that is a scandal. It only becomes one when the result hides it.
Consider a bakery owner checking a new sack of flour on the kitchen scale. If the scale says 24.6 kg for a 25 kg sack, the owner wants to know whether the scale is good to a gram or to half a kilo before complaining to the supplier. The reading alone does not answer that. Grain grading is the same question with more at stake.
How we are building the range in
In GrainGrade, every number is designed to be shown with its range, never as a bare decimal. The counting step makes this concrete. Rather than trusting one method, the pipeline produces three independent estimates of the grain count; the published count comes with a range, and the result is flagged if the estimates disagree. On our benchmark, counting error is measured at 3.5%, against a target of 3% or less. We publish both numbers because the gap between them is real.
| Bare result | Result with a range | What it lets you decide |
|---|---|---|
| Shrivelled & broken: 5.8% | 5.8% with its range shown | Whether the sample is clearly under the 6% FCI FAQ limit or too close to call |
| Count: 412 grains | 412, with a range from three estimates | Whether the estimates agreed or the photo needs retaking |
| Grade: FAQ | FAQ, with the defect ranges behind it | Whether a value cut is clearly justified or borderline |
The values in that table are illustrative, not measured results.
Ranges make borderline lots visible
The real value of a range shows up near a limit. If a sample’s shrivelled and broken percentage is well below the limit even at the top of its range, there is nothing to argue about. If the range straddles the limit, both sides can see the result is borderline, and the sensible step is another sample or a lab check, not a fight over the second decimal place. A bare number hides that distinction and turns honest uncertainty into a dispute.
Refusing a bad photo is part of the same idea
A range only means something if the input was good enough to measure. That is why GrainGrade is being built to refuse a bad photo with coaching rather than analyse it into a wrong answer. Six capture checks are planned before any grain is counted, including the phone within ±8° of level, a single layer of grains, a scale marker in frame and a check for blur and exposure. If the photo cannot give a trustworthy answer, the app asks for another one.
And every result should be re-derivable
Each analysis stores the exact settings and version used, so any result can be re-derived later. If a buyer and a seller disagree next month, they can see what was computed and how, not just a number on a slip.
For the limits these ranges are compared against, see FCI FAQ wheat specification explained, and for how image grading compares with grading by eye, manual wheat grading vs image analysis. The GrainGrade guides page has the rest. Grading and procurement terms are commercial decisions; confirm them with your own quality and legal advisers.
Questions
Why should a grain grade show a range?
Because every grading method has error. A range shows whether a result is clearly inside or outside a limit, or too close to call, which a bare decimal hides.
How accurate is GrainGrade’s counting?
GrainGrade is in development. Counting error is measured at 3.5% on Autowhat AI’s benchmark, against a target of 3% or less. Other figures are design targets.
What happens if the photo is poor?
GrainGrade is being built to refuse a bad photo with a reason in Hindi or English and ask for another, rather than produce a number it cannot trust.
See it on your own data. GrainGrade — Wheat graded from a photo. Book a 30-minute working session with an engineer.