AI developers struggle to trust model outputs when reasoning traces may be misleading or fabricated.
Build a library that ingests model outputs and traces, then analyzes them for consistency using techniques like counterfactual testing and temporal analysis.
Sell to AI teams at mid-sized companies implementing LLMs, who need audit capabilities but lack resources to build in-house.
MVP could be a Jupyter notebook plugin that flags suspicious reasoning patterns in existing traces.
Biggest risk is that the field moves toward different explanation paradigms, making trace analysis obsolete.