From this release, every model in the Terravek portfolio ships with a model card: the training data, the evaluation methodology, the measured performance, the known failure modes and the populations or conditions where it should not be used.
The failure modes section is the one that causes discussion. Our vegetation encroachment classifier, for instance, degrades substantially under dense evergreen canopy — recall drops from 0.94 to 0.61. That is written down, in public, in a document a competitor can read.
The argument for publishing it is straightforward. A customer using GRIDSCAN to prioritise maintenance on a network with significant evergreen cover needs to know that low-confidence spans are low-confidence because of the canopy, not because they are clear. Without that, the model is actively dangerous.
The argument against publishing it is that it makes our numbers look worse than a competitor who has not measured theirs. We have decided to live with that.
The model cards are in the product documentation and the methodology is described in the Terravek Labs technical note published alongside this post.