After this weekend, this is roughly how I expect a lot of my future non-implementation work to look. 1. Start with a frontier model and use it to really open up the problem. 2. Make those explanations explicit hypotheses. Then work out what would actually distinguish them. If two hypotheses imply the same observations, they are not yet useful. 3. Take those discriminating questions and run them against the raw evidence with a probabilistic semantic model like Jev. 4. Keep the result as state rather than turning it straight back into prose. 5. Look at what is still unresolved and ask which missing observation would do the most to separate the remaining hypotheses. The ontology underneath this is a graph. You have hypotheses, evidence, etc etc. interlinked. they store both relations and probabilities. Technically, though, if the graph contains all the information needed for the next update, then you can define the graph itself as the state and make the update process Markovian. This really isn't anything new to be fair. Hypothesis generation is basically an abductive problem and choosing the next observation is Bayesian experimental design. The key thing here is looking a set of tools so that they don't dogfood themselves, i.e. judging their own homework in the same transcript, which tends to bork the quality of reasoning. Very bullish on my future extended mind.