Content governance

Content governance applies exit conditions to the generated prose of a project: its documentation and its marketing. It splits along one line: rules that can be decided deterministically become checks the engine runs, and rules that need a judgment become delegated checks the AI assistant runs and records. The engine never calls a model.

Deterministic rules delegate to Vale

Most content rules are deterministic, and they become introspective auditors that delegate to the Vale prose linter, the same engine that already runs in CI. CONTENT-LANGUAGE covers this class. It is a prohibition (NEVER) condition — a banned construction must never appear, so an unaudited CONTENT-LANGUAGE reds rather than passing. It is a content-brand-governance condition and ships OFF by default; a project opts in.

CONTENT-LANGUAGE checks that:

  • a banned phrase is absent (Vale existence);

  • a mandated substitution is honored, for example "use" not "utilize" (Vale substitution);

  • reading grade is within a bound (Vale metric);

  • terminology is correct, for example the full product name "Akka Specify Plugin".

The engine shells out to vale --config=<class>/.vale.ini and maps a non-zero exit to red, the same way a declarative command auditor does. Structural rules — required sections present, heading conventions honored — are also deterministic and stay in the Documentation & Training type.

Tone is a delegated check

The remaining rules are holistic and cannot be reduced to a token list: state a fact rather than a flourish, stay declarative rather than persuasive, read like a person wrote it. CONTENT-TONE covers this class as a completion condition resolved by a delegated check. Like CONTENT-LANGUAGE, it is a content-brand-governance condition and ships OFF by default.

The judge is not a 1-to-5 score, which rubber-stamps. It is a set of single-criterion pass or fail predicates, framed adversarially ("find every place this reads like marketing"), run by a model from a different family than the one that generated the text, and converted to a pass or fail threshold. The judge runs in the AI assistant, not the CLI. The assistant applies the decomposed rubric and records the result with an attestation. The engine records that the judge ran and covered the current content signature; it does not score prose. Because the attestation is keyed to a content-plus-rubric signature, a rule change re-stales it and forces the judge to run again.

Completeness is the doc surface class

Completeness of documentation is not a content-governance check. It is the doc surface class of the coverage gate: an enumerable documentable thing with no page reds the single coverage gate.

Per-content-class scoping

Rules attach to a content class the way Vale scopes by path. Product docs use plain technical prose, marketing uses brand voice, and case studies state facts rather than flourishes. The policy attaches an exit-condition set to a class, so the same engine can cover a separate marketing repository where the rules exist but are not yet enforced.

Related