# Discovery That Earns Its Standing A tested component does not make an untested whole trustworthy. A beam may be strong, a joint reliable, and a support well designed. But assembling those parts does not establish that a bridge will carry its intended load. Their arrangement creates relationships that did not exist when the parts were tested separately. The bridge must face its own challenge. The same principle applies to computational discovery. A useful structure can become material for something more complex, but the new construction must establish its own **warrant**—the evidentiary basis for relying on it. This architecture organizes discovery around that principle: build candidate structures, test their consequential behavior, retain useful constructions, and use them to reach further. At every new level, the resulting whole must face an epistemic challenge in the domain where it is intended to operate. ## More Than Producing an Answer There is a difference between calculating an answer and establishing that the answer is useful enough to act upon. A forecasting system can produce a number. A diagnostic system can assign a probability. A language model can generate an explanation. None of those operations, by itself, establishes that the resulting action or claim is justified. The architecture therefore treats a candidate as more than a representation or predictor. It includes the operating rules that determine how the representation is used. In the market application, mapping parameters and trading thresholds belong to the same evolving package. The representation determines how market experience is organized; the thresholds help determine when the resulting assessment becomes actionable. Their value is evaluated together. The question is not simply, “Does this representation predict well?” It is, “What happens when this complete package operates?” ## Experience, Challenge, and Selection The market implementation represents multiple supporting price traces on spatial maps. Historical configurations accumulate separate evidence according to subsequent positive and negative outcomes in the target market. A declared interpretation procedure converts that evidence into an assessment, and the package’s operating rules govern trading behavior. Crucially, observations excluded from building those evidence maps are used to challenge the package. The resulting trading performance feeds evolutionary selection. Testing is not merely a final report on a finished design; it helps determine what gets built next. The object under selection is the trading trajectory, not a collection of isolated forecasting errors. An individually inaccurate forecast may have little consequence if the system does not act on it. An apparently accurate predictor may still support an intolerable trading strategy. The intended objective is operational viability: a package whose behavior remains useful through the environment it encounters. The fitness function determines how that objective is expressed, including which rewards, losses, adverse periods, and constraints matter. ## How Discovery Can Become Cumulative The larger possibility appears when useful discoveries become material for further construction. A discovered structure may provide a new distinction, a useful representation, or a reusable procedure. Its availability can expand the space of things the system can subsequently build. Discovery changes the resources available for further discovery. But reuse does not transfer unlimited authority. A component that works in one setting may behave differently when combined with other components or applied at a new scale. The higher-level package must therefore face its own domain-specific test. Earlier success makes construction possible; it does not make the new construction correct. This creates two complementary processes: - Construction expands capability. - Evidence governs the conditions under which that capability may be relied upon. The ambition is cumulative discovery without cumulative exemption from scrutiny. ## What Might Become Possible The transferable principle is not a particular number of traces, map layout, or search algorithm. It is the coupling of construction, operating rules, consequential testing, and reuse. Different domains would require different representations, evidence, and challenges. ### Medicine A candidate package might combine a representation of patient information with rules for recommending an investigation, requesting additional measurements, or deferring a decision. Its challenge would concern clinical consequences—not merely agreement with historical labels. A useful component for recognizing one pattern would not automatically justify a treatment recommendation assembled from it. The opportunity would be to discover bounded capabilities: useful for particular patients, under particular conditions, with explicit limits. Patient benefit, harm, and appropriate clinical evaluation would have to govern the challenge. ### Scientific Discovery Candidate structures might include equations, explanatory models, experimental procedures, or new ways of organizing observations. A useful representation could reveal a relationship that was difficult to see before. That relationship might then become a tool for proposing further experiments or constructing a broader model. Each extension would still have to meet evidence capable of contradicting it. The potential is not simply faster hypothesis generation. It is a constructive process in which discoveries alter what can be investigated next. ### Engineering and Control A system might discover local control strategies, then combine them into more capable controllers. A controller successful in one operating range would supply useful material, but a combined controller would need testing for interactions, transitions, and failures absent from the component evaluations. Here, the architecture could support growing complexity without treating successful parts as certification of the assembled system. ### Language Models and General Intelligence A language model could serve as a powerful constructor of candidates: proposing explanations, programs, plans, and combinations of existing capabilities. The surrounding architecture would distinguish proposal from authority. Generated structures would become dependable resources only to the extent that appropriate challenges established their usefulness. The longer-term possibility is a system that builds increasingly capable constructions while repeatedly testing the new wholes. That is a direction toward cumulative intelligence, not a claim that general intelligence follows automatically. ## The Challenge Must Be Worth Surviving Everything depends on the quality of the challenge. A system selected against an inadequate measure may become excellent at satisfying that measure without accomplishing the intended purpose. Repeated use of the same evaluation experience can also produce specialization to that experience. Independent evaluation remains necessary. First principles establish why tested parts cannot certify an untested whole. They do not establish that any particular test is sufficient. The architecture’s promise lies in making that obligation explicit: every increase in constructive reach creates a corresponding need for new evidence. Its deepest possibility is a discovery process that can build beyond what its designer could directly specify—while requiring each new construction to earn its standing in the world.