The corpus membrane

HLC calls the boundary between cultural records and a model’s training material the corpus membrane. The term groups together decisions about collection, access, selection, filtering, and preparation. It is useful because it interrupts the easy assumption that a language model has absorbed culture in an even or comprehensive way.

A membrane permits some material to pass and excludes or transforms other material. Here that is an analogy for a set of processes, not a literal component with one standardized implementation. The details differ across systems. Without documentation for a particular model, HLC vocabulary cannot tell us exactly which works were included, how they were weighted, or what was removed.

The digital trace therefore sets only an outer possibility. Even a surviving, publicly available document may be absent from a given training collection. Material that is present can still be hard to elicit accurately. Availability, learning, retrieval, and faithful reproduction are different questions.

For an exploratory session, this suggests two habits. Supply relevant sources when you have them, and distinguish claims based on those sources from broader model-generated background. Then ask which perspectives remain absent from the supplied collection. A carefully chosen packet of documents can improve the local basis of a discussion without making it representative of everyone affected.

The same distinction helps when comparing multiple models. Differences in output might reflect many factors, including training and later tuning; the comparison alone does not identify the cause. The membrane reminds us to keep causal explanations proportionate to the information actually available.