Perturbation atlases increasingly connect chemical or genetic interventions with high-dimensional cellular responses, yet their pharmaceutical value remains limited by a persistent reasoning gap. A perturbation signature records what changed under a particular experimental condition, but it does not independently establish why the change occurred, whether it is reproducible across biological contexts, whether it reflects therapeutically achievable exposure, or whether it supports target validation, repurposing, or combination design. This Original Perturbation-Informatics Architecture Article proposes the perturbation reasoning engine as a conceptual architecture for transforming perturbation observations into qualified, context-dependent pharmacological hypotheses. The proposed construct treats each observation as a structured relation among an intervention, dose or perturbation strength, exposure duration, baseline cell state, biological model, assay context, response distribution, and provenance. It then separates representation from subsequent reasoning functions, including directional inference, mechanism attribution, interaction analysis, resistance interpretation, cross-platform harmonization, uncertainty qualification, and bounded decision use. The central contribution is an information architecture in which no single predictive or similarity metric is treated as sufficient evidence of pharmaceutical usefulness. Instead, claims are qualified according to their evidentiary burden, context dependence, transportability, and susceptibility to alternative explanations. The architecture is intended to organize theory building, benchmark design, experimental prioritization, and evidence reporting rather than to constitute a validated predictive system. Its principal limitations include incomplete perturbation coverage, uncertain intracellular exposure, non-equivalence between genetic and pharmacological interventions, latent-state ambiguity, platform shift, and the absence of prospective confirmation for many computationally inferred relationships. Perturbation atlases may therefore become more useful as pharmacological reasoning resources when they preserve context, expose uncertainty, distinguish association from mechanism, and return computational hypotheses to discriminating experiments.