A run used to be a fixed script. Now it stops at every point where budget is spent and asks Claude: after each search, which authors deserve a probe; after the probes, which accounts deserve a full measurement and whether they are on topic, adjacent or off. The rubric Claude judges with is written down and updated as the run learns.
Every collection keeps a memory: the objective, the queries that produced winners and the ones that were retired, the known winners and why they win. The next run in the same niche starts from that instead of from zero.
- Triage, probe and verdict decisions made by Claude, with a brief at each step
- Cheap probes before expensive measurements
- Collection memory: objective, productive queries, retired queries, known winners