been working to integrate this into my harness for the last few hours. So far the best thing I’ve found is parsing blocks of text to find hidden obligations dropped by llms. It’s about 75% accurate which is about 10,000x more accurate than Astra above 300k context 😅