This paper develops a computational architecture for scientific discovery within bounded representational spaces. Because combinatorial growth makes exhaustive search impractical, the complete representational domain is distinguished from an effective search space shaped by scientific knowledge, empirical data, representation, and computational constraints. A constraint-first architecture is proposed in which independently justified scientific constraints guide candidate search while preserving the distinction between candidate identification and scientific validation. A retrospective rediscovery experiment provides a computational proof of concept. Dimensional consistency reduces a canonical domain of 29,524 symbolic candidate relations to nine while retaining a withheld ideal-gas relation, a 99.96952% reduction. Empirical invariant scoring ranks the target first in all 800 evaluations across four tested conditions, including the noise-free condition. In a full-space baseline using the same scoring procedure, both methods rank the target first in all noise-free evaluations; under the three noisy conditions, constraint-first search ranks it first in all 600 evaluations, whereas full-space ranking does so in none. Thus, in this controlled benchmark, independently specified physical constraints both reduce empirical search and improve robustness to noise. The experiment is deliberately restricted and does not establish general-purpose autonomous scientific discovery. ( direct link )