The Siemens challenge dataset at Nebius Build Week has 10,000 anonymized suppliers. Each one has 18 risk scores from 0 (no risk) to 100 (maximum risk) and a Total Risk Index, described only as "a weighted average of all risk categories". The weights were not given. We recovered them exactly. Before you look, try to guess.
1 / Guess
Which three scores weigh the most?
Pick three of the 18. Three of them share the top weight, so there is one right answer.
0 of 3 picked
2 / Reveal
The 18 weights
The formula
Round half to even means a sum that ends in exactly .5 goes to the nearest even whole number: 50.5 becomes 50, 51.5 becomes 52. Every other value rounds to the nearest whole number as usual.
Reproduces 10,000 of 10,000 rows (round half to even).
10,000 / 10,000rows reproduced exactly with whole percent weights
0.9996R² of the least squares fit
0.25mean absolute error in points, before rounding
−0.017fitted intercept, so no constant term
3 / Play
Build a supplier
A synthetic supplier, not a row from the dataset. Move any score; the index is recomputed with whole number arithmetic and round half to even, exactly as for the real rows.
50Highweighted sum 50.00
Low 0Medium 25High 50Critical 75
Smallest single change that flips the level
Lower level
Higher level
Each row shows the score, then its points in the index (weight times score). Points always add up to the weighted sum before rounding.
4 / Method
How the weights were recovered
Least squares. We regressed the published index on the 18 scores over all 10,000 suppliers, with an intercept. The fit explains the index almost fully (R² 0.9996). The intercept came out at −0.017, close enough to zero to drop, and no weight was negative.
Round to whole percent. Every fitted weight sat within about 0.0003 of a two decimal value (0.0999 for 10 percent, 0.0301 for 3 percent), and those values add up to exactly 100 percent. The leftover error, 0.25 points on average and never more than 0.53, is the size you get from rounding the index to a whole number.
Exact check. With the whole percent weights, the weighted sum is a whole number of hundredths, so it can be computed with no floating point error. Rounding it half to even matches all 10,000 published values. Rounding half up matches 99.53 percent: 96 rows land exactly on .5, and the published index sends them to the even neighbour. Ordinary floating point rounding matches 99.92 percent, because a few of those ties drift by a tiny error.
Why it matters. A buyer can now say "Operational Resilience 95 adds 9.5 points to this supplier's index" as a fact, not an estimate. One point on Operational Resilience, Financial Health or Quality Risk moves the index as much as 3.3 points on Media Sentiment, Market Sentiment or External Intelligence.
Aggregate numbers from the dataset
Measure
Value
Suppliers
10,000
Total Risk Index range, mean
6 to 93, 50.15
Suppliers per level (Low, Medium, High, Critical)
318, 4,485, 4,813, 384
Same 18 scores with equal weights: R², exact after rounding
0.983, 21.9%
Recovered weights: R², exact after rounding
0.9996, 100%
Index below 50 but at least one Critical score
911
Level bands are the same for every score: Low 0 to 24, Medium 25 to 49, High 50 to 74, Critical 75 to 100. The weights are also in weights.json. No supplier rows are published here.
Part of Supplier Risk Radar, a private project by Ciprian Marian, started at Nebius Build Week in Berlin.