Abstract

Autonomous discovery loops, in which language-model agents improve a design against an automatic evaluator, spend most of their effort on ideas that fail. We argue that such a loop should make failure cheap, leave success uncapped, and score the part that decides what to try like everything else. We introduce recursive priced falsification, an agentic topology: an arrangement of roles, what each may read, the gates between them and a shared record. Instruments measure the incumbent design, scouts write priced cards, a pricer runs the cheapest test that could end a card, a builder builds what survives, and a breaker in a fresh context attacks each claim. Every task is a trial: a forecast written before the run, the result that would end the idea, and its price in the objective’s currency, so the stake is small and known while the gain is uncapped. Trials are ranked by expected gain per unit of stake. The method is recursive: the loop arrived at its own roles by these tests, the planner’s forecasts are scored, a red team attacks closed directions, and the loop revises its procedure from its mistakes. We give optimality conditions for the price map and ranking rule, an order for an evaluation cascade, and a rule for exactness. Over six days on a live public benchmark for quantum circuit optimisation, beside other agent harnesses, the loop produced 70.8% of a 5.73% fall in the best score with 17 of 42 accepted entries, including the only two steps above 1%, while the benchmark ran twelve times its prior pace. The median trial cost 50 worker-minutes and kept nothing; ten of 133 trials hold 79% of the loop’s points. Trials beat their stated odds (46% against 30%), and the largest gains came from trials at about even odds, not long shots. We present this as an autonomous research loop, not recursive self-improvement: the model is frozen and the comparison is observational, with no ablation.

A preprint, not yet peer reviewed. 71 pages, 11 figures, 37 tables. The method is released as a skill and tools for coding agents at github.com/pensievelab/rpf.

Cite this paper

@misc{ghosh2026rpf,
  title        = {Recursive Priced Falsification: Asymmetric Trials with a
                  Forecast-Scored Planner for Autonomous Discovery},
  author       = {Ghosh, Krish},
  year         = {2026},
  month        = {October},
  howpublished = {\url{https://www.pensievelabs.co/papers/recursive-priced-falsification/}},
  note         = {Pensieve Labs}
}