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 and declares whether “yes” selects its primary anchor or an alternative. agree predicts yes, disagree predicts no, and pass means the faction takes no position. Stances are composed from issue-independent faction values and an issue-specific value trade-off, then oriented to that declared polarity. Canonical question wording remains authoring metadata, not a hidden lexical requirement in environment scoring.

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 and v6 template/layout-profile labels are disjoint. The 0.6 candidate contains 100 training and 100 held-out rows. The generator assigns opaque document/faction IDs, varies visible layout and faction count/order, prints round-trip-exact faction values in prompt-visible summaries, and records question polarity, continuous preference-trade-off value, and explicitly authored paired evidence for every contradiction. Each planted issue also carries a hand-authored five-slot decision reference for Ask. Find intentionally omits that reference and scores visible evidence, issue type, valid yes/no diagnosis form, and contradiction relationships. Ask renders the canonical five-slot frame and exact signed trade-off weights for every unordered candidate while retaining evidence keys, issue labels, relationships, stored stances, decision values, and utilities as answer metadata. Validation rejects authored relationships that collide with another planted issue or distractor.

Split generation distinguishes exact-instance and policy-issue fingerprints; it also runs one word-ngram TF-IDF cross-split neighbor audit. Generated scenarios are canonical JSON and regenerate byte-for-byte from the same template, seed, and explicit generation date.

The 18 helper-built held-out families separate each primary issue, authored contradiction relationship, and plant-free ledger into five documents. Authored policy sentences remain intact. Seeded type-neutral administrative distractors are composed separately before sentence order, title, and style randomization. Each procedural predicate is used at most once within a scenario, preventing filler from inventing repeated cross-document responsibilities. Balanced filler counts and seeded sentence shuffling keep issue position from becoming a fixed first-clause rule. Historical reports retain earlier shortcut audits; the active gate keeps compact prompt-only probes and a top-one utility diagnostic for the default K=1 Ask budget. Bounded views retain only issues whose authored anchor remains visible.

The bundled 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:5, requires persona_panel.vote_rule: value-composition-v2, and is available through load_scenario_schema().