A negative result on every tile

Over the last stretch the Faena hub went from three live mining tools to eleven. Writing that sentence, the easy move is to talk about coverage: exploration, drill and blast, load and haul, comminution, geotechnics, economics. The more useful thing to say is about a rule I gave myself while building them, and what it changed.

The rule is that each tool has to be honest about where its learned model does not help. Not in a footnote. On the tile, in the benchmark, where a user looking for a reason to trust the tool would land.

In practice that turned into a negative result per tool, and they are more interesting than the wins:

• In TailWatch, an InSAR ground-deformation studio, I trained an anomaly autoencoder to flag accelerating slopes. On the held-out benchmark the classical velocity map (AUC 0.968) beats it (0.898). The learned model is the loser, and the page says so.
• In DispatchLab, a truck-shovel dispatch bench, the policy I most wanted to work was a distilled Monte-Carlo rollout under cycle-time uncertainty. It wins 0 of 8 random seeds against a simpler assignment. That is a published null, not a hidden one.
• In PitForge, the exact optimizer (a min-cut solving the ultimate pit limit to within 2e-9 of three published MineLib optima) is the star; the grade neural net ties ordinary kriging (0.9613 vs 0.958) and never beats it.
• In ProspectMap, on a real mineral belt, a trivial distance-to-deposit null (0.783) ties the best model. Most of the apparent skill is spatial proximity, and the app commits that verdict as ranking_win: false.
• Even ChronoScope, a forecasting atlas, keeps a case where a one-line SeasonalNaive beats a foundation model on real data.

None of this is self-deprecation. Each tool has a real, exact or classical core doing the work, and the learned layer is measured against it in public. What the rule bought me is trust that survives contact: an operator who finds the one place the tool is honest about its own failure has a reason to believe the rest. A single flattering accuracy number buys the opposite, and it is the number most tools show.

There is a version of a portfolio that is a list of things that worked. This is the other kind: a list of things I measured, including when the fashionable method lost to the boring one. I think it is the more valuable list to have built.