Rashomon Set
Manager
Any serious model team ends up with a set of models that all look equally good on validation. Pick one arbitrarily, and you've silently accepted a risk profile you never evaluated. The Rashomon Set Manager treats the ensemble as the unit of deployment — registered, profiled, health-checked, and continuously revalidated.
Dashed polygon: target coverage. Solid: current pool. Adversarial axis underperforming — consider adding a robust model.
Your ensemble,
governed as one object.
Register models from PyTorch, TensorFlow, JAX, scikit-learn, ONNX Runtime, Hugging Face, SageMaker, Vertex AI, or Azure ML. The manager continuously profiles each one, labels its architectural family, and tracks how its reliability decays across forecast horizons.
- Multi-framework registration via ONNX or native adapters
- Automated reliability-decay profiling per model
- Architecture family tagging (Transformer, Reservoir, CNN, LLM, GBDT…)
- Diversity scoring and blind-spot band detection
- Drift-triggered revalidation and scheduled health checks
- Pool versioning with rollback
from Rebound import Pool
pool = Pool.create(
name="wind_dispatch",
risk="balancing_penalty",
)
pool.register("transformer-v3", onnx="s3://models/tf_v3.onnx")
pool.register("reservoir-net", torch="models/res_net.pt")
pool.register("cnn-lstm", torch="models/cnn_lstm.pt")
pool.register("gbdt-baseline", joblib="models/gbdt.pkl")
pool.register("llm-finetuned-b", hf="org/llm-ft-b")
report = pool.profile(
dataset="wind_2025_q4",
horizons=[1, 6, 12, 24, 36, 48],
)
report.diversity # 0.81
report.blind_spots # ["k > 36h"]
report.recommend # "add: adversarially-robust TCN"Already have 3+ models in production?
That is a Rashomon set. We will characterize it for you.