Update: the investigation loop initially hit its 4-step budget without
voluntarily concluding (documented honestly rather than hidden). Fixed with
a forced-conclusion step -- when the budget runs out, one final call with no
tool options asks the model to summarize only what it actually observed.
Every investigation now ends in a real, grounded answer, including ones that
honestly say the evidence wasn't conclusive rather than overclaiming. Also
added make investigate -- run the agentic loop in isolation, one command,
no need to read through the full pipeline's output.
Log in or sign up for Devpost to join the conversation.