VariantCut

VariantCut tells a car maker which versions of a model to stop selling: the powertrain variants that cost around €30M each over their life but sell only a few hundred cars a year. Today these calls are mostly made on volume alone. A team of Muse Spark agents works only on public data (Dutch RDW registrations with list prices, and the EEA CO2 monitoring dataset, 2019–2025): • Analyst scores every version through as-of tools on anonymised data (hashed names, power and price rounded), so it can't see the future or recognise the car. • Critic checks each proposed cut against binding rules (the group's fleet CO2 budget, last version of a fuel, cheapest way into the model, still ramping up) and debates the Analyst over three rounds. • Substitution, learned from 162 Dutch discontinuations, says where buyers go. Finance prices every cut with a cited cost model and the EU CO2 penalty. • Muse Spark writes a one-page board memo. Tested like a forecaster: calls made with 2022 data, checked against what makers really dropped by 2025. Cutting the smallest versions is right 40% of the time; the full team 52%. On held-out years (2019 → 2022) it is 68% vs 57%. A recognition probe shows the edge is larger on cars Muse Spark could not name: the gain comes from reasoning, not memory. A Live view applies the team to 2025 data, with an early check against 2026 Dutch registrations. Demo: https://rahulthms-art.github.io/VariantCut/