Bendex opens training-stability core

This is a strong infrastructure pitch if the results hold outside the demo: catching divergence early is far more useful than reacting after loss explodes. The open-core angle also makes sense here, because training stability is the kind of problem teams will trial in research and then pay to operationalize.

• –The differentiator is geometric monitoring of weight updates, not another loss or gradient dashboard
• –The site claims 100% detection, 0% false positives, and 90% recovery on a 30-seed benchmark, which is impressive but still self-reported
• –Real value is likely in preventing wasted runs, especially for expensive fine-tuning and large-scale training jobs
• –The product already looks like an open-core funnel: free research version, paid Pro and Enterprise tiers for intervention and licensing
• –If the attribution and intervention logic generalizes, this could slot into MLOps stacks alongside checkpointing, alerting, and training orchestration