Who or what should we ask next?
Pukot's current product is practical: define bounded questions, gather independent judgments, measure quality, retain useful disagreement and deliver structured evidence. Our longer-term research asks how routing can improve from demonstrated usefulness.
Current capability
Bounded tasks
Questions, criteria and answer spaces are defined before collection.
Independent judgment
Separate responses can be retained so early consensus is not automatically treated as truth.
Project-specific quality
Quality controls and provenance are designed around the program rather than a universal score.
Research direction, not deployed claim.
We are exploring whether a distributed network of human and synthetic contributors can progressively learn which contributor is useful for which class of question. The question is not who is smartest in general. It is who or what produces useful evidence under these conditions.
Status, credential, geography, model confidence or contributor type are not sufficient proxies for usefulness. The evidence should determine what gets asked next.
Methods we care about.
Calibration
Where forecasts are resolvable, confidence should be compared with actual outcomes rather than agreement with another forecaster.
Disagreement analysis
Divergence may identify ambiguity, preference, missing context, rubric problems or genuine heterogeneity.
Task-conditional capability
Contributor value should be understood by task and domain rather than collapsed into one permanent rank.