- Mass balance and reconciliation
- Assays and lab data
- Agents, tools and skills
- Language models
- Tailings and water
Can a local AI assistant tell which size fraction carries the most copper?
Published table · one tailings sample · two local models with Python · 20 attempts
We gave two local AI models, each able to run Python, a published table of copper assays for seven size fractions of one copper tailings sample, and asked which fraction carries the most copper. Ten attempts each; we checked every reply against our own sums of mass times grade.
What we found
- Right fraction, every timeAll 20 attempts named the finest fraction: a third of the sample's mass at 1,084 ppm Cu, 31 % of the copper the table reports.
- What the missing mass could changeThe table's masses add to 91.8 %. 9 replies mentioned it; 0 of 20 said the missing 8.2 % might hold copper too. If it all belonged to the coarsest fraction, at that fraction's grade, the coarsest would carry the most.
- Richest is not mostThe coarsest fraction has the highest grade, 2,250 ppm, but only 8.6 % of the mass: 17 % of the copper the table reports.
- Right ranking, unreliable detailsAll 20 scripts put the fractions in the right order. 11 of 20 replies then stated a wrong number or comparison.
Method
One published table, two models, twenty attempts
- RealTable 1 of Habibi et al. (2025)Seven size fractions of one sample of fresh flotation tailings from the Carrapateena processing plant, July 2024. Two separate readings of the paper agree on every value.
- TestedTwo local models with PythonQwen 3.8 27B and Qwen 3.8 Flash Next on our own GPUs, each in a coding agent with Python and no web search, ten attempts each with the table and the bare question.
- Checked by handEvery replyRead against our own sums. Twenty attempts on one table show what these two models did here, not how they do on every table.
The prompt, the run scripts, every reply and our reading are in the article's public repository: code under the MIT licence, results under CC BY 4.0. The table is reproduced from the paper under its CC BY 4.0 licence.
- Habibi Z., Cook N.J., Ehrig K., Ciobanu C.L., Campo-Rodriguez Y.T., King S.A. (2025). Minor and Trace Elements in Copper Tailings: A Mineralogical and Geometallurgical Approach to Identify and Evaluate New Opportunities. Minerals 15(10), 1018. doi.org/10.3390/min15101018
01 · Our check
The richest fraction is not the one with the most copper
Copper carried is grade times mass. CTS+106 has the highest grade but is a thin slice of the sample; CTS−C5 is a third of it. The masses add to 91.8 %: choose where the missing 8.2 % belongs and watch the answer move.
Length: the copper grade. Height: the share of the sample's mass. Area: the copper the fraction carries. “Part of” a fraction counts the missing mass at that fraction's grade.
As the table stands, CTS−C5 carries 31.1 % of the copper it reports, nearly twice CTS+106's 16.7 %. The paper does not say what the missing 8.2 % is.
The answer can change
If the missing mass was all CTS+106, that fraction would carry the most copper. As a fraction of its own it would need above 4,402 ppm, twice the richest grade.
So can the share
31.1 % is CTS−C5's share of the copper the table reports. Across the readings in the figure, up to the richest grade, its share of the whole sample runs from 27 to 36 %.
02 · Twenty attempts
CTS−C5 every time; the missing copper never
Each model got the table's three blocks of numbers and the bare question ten times, in a coding agent with a shell, so it could run Python. Every attempt ran a script. We read every reply against our own sums. Pick an attempt to see what it said.
The 27B saw it and said nothing
In two attempts the 91.8 % showed up in its own script output or its thinking. “No problem. The ranking is valid,” it thought, and left it out of the reply.
Flash Next said more, not all of it right
Nine Flash Next replies flagged the 91.8 %, as a reason to rescale the shares, then carried on into minerals and recovery. Three gave the gap a cause, such as “processing loss”, that the paper never states.
03 · What the table cannot say
Sizes, minerals and a missing head grade
A careful answer says what the table leaves out before it answers. Three things this one leaves out, and what the replies did with them.
Read as “5 µm”, the code C5 became a size. The paper gives none for C1, C3 or C5; its abstract puts the finest fraction at −10 µm.
What the replies said about the copper's minerals
- Liberated in the coarse grainsFlash Next 5, 7, 8
- Locked in, likely refractoryFlash Next 4, 6
- With silicate gangueFlash Next 2, 3, 4
- Consistent with fine chalcopyrite27B 3, 9
- Where it can be recoveredFlash Next 2 to 10
One table, five stories. Assays say how much copper each fraction holds, not which mineral holds it or whether it can be recovered. For that the paper's authors turned to mineralogy.
No head grade to check against
The paper reports no measured head grade for this sample. The 1,263 ppm several replies quoted is the table's own average, so it cannot show whether the missing 8.2 % carries copper.
One sample
Fresh flotation tailings, sampled in July 2024, one reported assay result per fraction. The table describes that sample, not every day's tailings.
04 · Before you trust it
Before an assistant reads your site's lab tables
Four questions for any reply, from a person or a model.
- Grade, or copper carried?Name the one you mean.
- Does the table close?Masses to 100 %, and a head assay the fractions add back to.
- Is each number in the words right?Not only in the script.
- Where did each size, mineral or cause come from?The source, or a guess?
Both models ranked the fractions right. Neither explained how the missing 8.2 % could change the copper accounting. Check that part by hand.