Exposure modes

The exposure mode controls which parts of the hidden world are written into the output bundle. This is a first-class design decision: truth is filtered during rendering, not by omitting simulation.

Mode	CLI flag	Who it's for
student_public	--mode student_public	Teaching, portfolio projects, Kaggle-style competitions
research_instructor	--mode research_instructor	Instructors, researchers, break-me analysis

student_public

The default. Includes:

• All 9 relational tables — but with post-conversion entities omitted and timestamps snapshot-capped
• Task splits (train.parquet, valid.parquet, test.parquet)
• feature_dictionary.csv and dataset_card.md
• manifest.json (provenance without hidden-world details)

Excludes:

• Latent registry (trait values)
• World spec (DGP parameters, recipe instantiation)
• Hidden causal graph (graph.graphml, graph.json)
• Mechanism summary
• customers and subscriptions tables (their existence reconstructs the label)

research_instructor

Everything in student_public, plus:

• metadata/world_spec.json — full DGP and recipe instantiation
• metadata/graph.{graphml,json} — the hidden causal DAG
• metadata/latent_registry.json — per-entity latent trait values
• metadata/mechanism_summary.json — per-edge mechanism parameters
• Full-horizon relational tables including customers and subscriptions

The instructor companion dataset on HuggingFace (leadforge/leadforge-lead-scoring-v1-instructor) was built in this mode.

Monotonicity guarantee

The exposure modes satisfy a monotonicity invariant: every artifact in student_public is also present in research_instructor. You can always load the student view from an instructor bundle; you cannot load the instructor view from a student bundle.