17.6 Why a corpus is not a simulator

A simulator executes a declared transition. A document reports, interprets, or disputes one. This difference changes what can be admitted. Replacing a mechanism in code may generate a counterfactual trajectory; replacing a phrase in a paper generates only another phrase. Corpus-to-theory construction must therefore separate four operations that a fluent summary can easily collapse:

1. extracting what a source claims;

2. typing the context and evidence supporting that claim;

3. deriving consequences of several claims under a proposed theory; and

4. testing those consequences against evidence not used to construct the theory.

Even the fourth operation is corpus-relative when its evidence is another paper. Publication bias, shared datasets, copied assumptions, and common measurement errors can make nominally independent sources dependent. A locked document can test prediction and provenance without identifying a natural mechanism.

There are also several meanings of absence. A relation may be false, unmeasured, unreported, excluded by the search query, or lost by the extractor. A source-bearing sketch records these possibilities separately. In particular, the absence of an edge is not automatically a negative edge, and a failed global gluing operation is not automatically a latent biological variable.