Recursive Priced Falsification: Asymmetric Trials with a Forecast-Scored Planner for Autonomous Discovery

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 be built so that a failure is cheap and a success is not capped, and that the part that decides what to try should be scored like everything else. We introduce recursive priced falsification, an agentic topology that does both. It is an arrangement of agent roles, what each role may read, the gates between them and a record they share. Instruments measure the incumbent design, scouts write priced cards from that map, a pricer runs the cheapest test that could end a card, a builder builds what survives, and a breaker with a fresh context tries to break each claim. What moves between the roles is a trial: a forecast written before a run, the result that would end the idea, and the idea's price in the currency of the objective. Each trial is an asymmetric bet: a stake that is small and known before the run, against a gain that the stake does not cap. Trials are ranked by expected gain per unit of stake, not by their chance of success. The method is recursive: the loop applies it to itself. The loop arrived at its own arrangement of roles this way during the run: it began as a hub, priced candidate roles against its own record, and adopted the arrangement after running it once as a trial, with a forecast written first. The planner that ranks the trials is forecast-scored: it commits to a forecast in every task, every report lists what the task got wrong, a red team attacks its closed directions, and the loop revises its own procedure from its logged mistakes. We give the optimality conditions behind the price map and the ranking rule, an ordering rule for an evaluation cascade, and a rule for which comparison can decide that a change is exact. We implemented the topology in Pensieve, our research harness, and ran it for six days on a live public benchmark for quantum circuit optimisation, with other agent harnesses improving the same circuit under the same evaluator. The best score fell by 5.73% in 42 accepted entries. The loop produced 70.8% of that fall in 17 entries, took the top of the benchmark twelve times, and made the only two steps above 1% by any system (1.45% and 1.04%). The benchmark as a whole moved twelve times as fast during the run as in the five weeks before it (0.99% a day against 0.083%), and fell back to 0.063% a day when the loop stopped. Its trials paid in the way the design predicts. The median trial cost 50 minutes of one worker and kept nothing; the mean trial earned 0.03% of the score, the size of the median accepted entry of the other systems. Ten of 133 trials hold 79% of the loop's points, and the largest of them cost one builder 63 minutes. Trials beat the odds written for them (46% against 30%) and returned 1.6 times the size forecast for them. What paid was a capped loss against an open gain: the largest gains came from trials given about even odds, not from long shots. The loop revised its own procedure during the run, after each of 67 logged mistakes; after the full topology was in place, 1 of 12 builds kept nothing, against 10 of 35 before. We present it as an autonomous research loop and not as recursive self-improvement: the foundation model is frozen, and no revised loop was tested against its predecessor under a matched budget. The comparison with the other systems is observational and has no ablation. All research was carried out by the agents; a human operator set the goal, approved actions with effects outside the loop and sent occasional directions on strategy.