Difficulty profiles

Each recipe ships with difficulty profiles that modulate how hard the prediction task is. The profiles adjust noise, missingness, and signal strength — not the underlying causal structure.

Profile	AUC (approx.)	Conversion rate	Who it's for
intro	≈ 0.89	~28%	First-time learners, pipeline sanity-checks
intermediate	≈ 0.79	~18%	Standard benchmarks, course projects
advanced	≈ 0.68	~8%	Experienced practitioners, calibration and rare-event work

Set via CLI

leadforge generate --difficulty intro    # or intermediate / advanced

Set via API

bundle = gen.generate(n_leads=5000, difficulty="advanced")

What changes between tiers

• Noise level: measurement noise on engagement features, channel attribution, and timing
• Missingness: the advanced tier has higher rates of missing values in several feature categories
• Signal strength: the coefficient magnitudes on causal paths are attenuated in harder tiers
• Conversion rate: directly calibrated so each tier hits its declared band

What does not change: the motif family, the graph structure, the entity IDs, the recipe narrative.

Profile YAML

Difficulty profiles are defined in the recipe:

leadforge/recipes/b2b_saas_procurement_v1/difficulty_profiles.yaml

You can inspect the full calibration parameters by reading this file or by running:

leadforge list-recipes --verbose