Why “Native Semantic Form” Matters – We Use The LLM’s Grammar, We Don’t ‘math it’.
LLM producers often think: “If it’s serious, it belongs in a database with schemas.”
But natural langauge has a schema. We just narrow it into operational prose.
So our strategy is different: we exploit that most institutional knowledge already exists as semantically structured text:
-
policies, contracts, statutes, guidelines, SOPs
-
case narratives, incident reports, clinical notes
-
argumentation, exceptions, defeaters, precedence
-
definitions and scope conditions
Relational databases excel at extensional facts (rows/columns). They are poor at intensional structure (exceptions, precedence, defeaters, conditional obligations, scope clauses), unless you re-encode everything into a bespoke logic layer.
Runcible’s strategy is:
-
Keep normative/semantic artifacts in their native linguistic structure.
-
Compile them into tests and constraints rather than flattening them into relational calculus.
-
Use the LLM as a semantic compiler that can map text into claim graphs + proof obligations.
-
Use the governance layer to force typed closure and prevent rhetorical completion.
This is the key “why it works” that labs miss: we are not askinging the model to “be moral”; we are using it to compile institutional semantics into computable checks.
Apparenly our use of morality and truth is confusing. Except, all language that is of value to humans that can be used by machines is in fact either both truthful, ethical-moral, possible, and liable or it isn’t.
So the foundation of everything … is ethics. Yes. Really.
So we start with ethics and build a governance layer.
That way we ‘cleans’ the world model of everything that isn’t true, ethical, moral, possible, and liable.
For some reason that set of ideas seems counter-intuitive to people – even people in the field.
Source date (UTC): 2025-12-31 19:17:28 UTC
Original post: https://x.com/i/articles/2006444612521713737
Leave a Reply