17 From Scientific Documents to Testable Theories

Corpus-Relative Construction and External Admission

Chapter 16 began with an executable world. A simulator exposed state variables, accepted interventions, and returned consequences against which a proposed theory could be tested. This chapter removes that privilege. Its input is a collection of scientific reports: papers, tables, databases, and extracted claims written by investigators who performed different measurements under different conditions.

Consider a hypothetical literature in which one set of studies reports that an intervention increases an outcome in one tissue, another reports a decrease in a second tissue, and a third proposes a mediator. The scientific problem is not to average the signs. It is to decide whether the reports describe a real context dependence, different populations or measurements, an omitted mechanism, or simply an extraction error. The desired output is therefore a testable theory proposal: a typed account of the entities and contexts, an exact ledger of supporting and opposing sources, consequences not used to construct the account, and an experiment capable of distinguishing it from its alternatives.

This changes the meaning of success. Documents can support reconstruction, comparison, and hypothesis generation, but they ordinarily cannot execute a new intervention. A corpus-relative result may show that a proposed structure explains several reports and predicts a sealed report. It does not by itself show that the structure is causally true of nature.

¹1. Chapter roadmap. Corpus-to-Theory Extension (CTTE) compiles controlled reports into source-bearing theories under increasingly hidden registration. GLP1–PILOT tests that spine in a scoped biomedical setting; GLP1–GRANT adds an auditable prospective program. These are progressively stronger obligations, not independent demonstrations of autonomous science. ↩

The reported evidence reaches typed transport in generated document worlds, bounded context construction in a curated GLP-1 interface, and conservative realization of a typed research program. It does not reach prospective biomedical admission, independent peer-review success, or autonomous grant submission. The long realization sequence later in the chapter audits how a source-bearing theory can cross into human-readable prose without acquiring unsupported claims; it is not an exercise in optimizing style.