---

name: sbi_simulator_prior_protocol description: Validate simulator, prior, observation, and simulation-record contracts for simulation-based inference recovery workflows.

SBI Simulator Prior Protocol

Use this skill when a recovery or implementation task needs to turn black-box simulator outputs into clean training records for simulation-based inference. It is appropriate for sbi-style workflows where parameters are sampled from a prior, passed to a simulator, and paired with observations for a neural density estimator or classifier.

Do not use this skill to train an estimator or to sample from a posterior. Its output is validated data and metadata for downstream SBI training.

Inputs

• Prior samples or a prior-grid specification with explicit parameter dimension.
• A simulator callable, simulator formula, or precomputed simulator outputs.
• Optional observation x_o and expected observation dimension.
• Optional records with simulator status values.

Outputs

• Validated records with theta, x, and status.
• Metadata with theta_dim, x_dim, num_requested, num_valid, num_failed, and shape warnings.
• A training dataset containing only valid numeric records.

Workflow

1. Normalize scalar parameters and observations to one-element float vectors.
2. Execute or ingest simulator outputs and build one record per parameter value.
3. Reject missing, nonnumeric, nonfinite, or dimension-inconsistent observations.
4. Preserve invalid records in the audit log while excluding them from training data.
5. Pass the valid (theta, x) records to an SBI training skill.

Validation

Run the deterministic tests:

python /share/project/yuyang/workspace/Paper2Skills/Distiller/skills/module-to-skill/scripts/validate_skill_tree.py /share/project/yuyang/workspace/Paperbench/record/case15/extracted_skills_attempt_001/sbi_toolkit/sbi_simulator_prior_protocol --run-tests

The included script can also be run directly on a tiny built-in example:

python scripts/simulator_protocol.py --demo

Limitations

This skill does not implement sbi internals, PyTorch tensor conversion, joblib scheduling, or neural-network standardization. It preserves the paper mechanism at the data-boundary level and leaves estimator training and posterior APIs to downstream skills.