commonground-scenarios

Deterministic, offline scenario generation for the Common Ground environment family. Each generated scenario contains a fictional organization, policy documents, planted ambiguities/contradictions/gaps, precise distractors, a seeded persona panel, and provenance.

Every canonical clarifying question is phrased as a yes/no proposition. Its target_stances have an explicit interpretation: agree predicts yes, disagree predicts no, and pass means the faction takes no position. Templates may also author a finite canonical_question_aliases allowlist. Aliases obey the same yes/no form and are kept distinct across the answer key.

The default generator uses committed templates only. An operator may inject a prose-polishing callback, but no model client or network path is built into the package, and validation requires every planted anchor to survive polishing.

Training and held-out template registries are disjoint. The canonical release contains four training templates with ten deterministic operative-scope contexts each and twenty separately authored held-out domains. A semantic fingerprint includes documents, planted anchors/types, questions, aliases, and faction stances while ignoring seed, organization name, and document order; split generation fails if two rows share a fingerprint. Generated scenarios are canonical JSON and can be regenerated byte-for-byte from the same template, seed, and explicit generation date.

The bundled 0.2.x scenarios are synthetic. The optional human_feedback field supports advanced custom input and delegates to the shared fail-closed validator. It is not used by the bundled data; see human-data governance for its separate review requirements.

The packaged Draft 2020-12 schema has the stable identifier urn:commonground:schema:scenario:1 and is available through load_scenario_schema.