ModelVariant
A trained model, identified by the horizon its states were assembled at.
Two per hotel, differing in training window and nothing else — same features, phases, policy class, hyperparameters, guardrails and masking. Further divergence should be earned from the shadow data, not assumed.
The value is a storage key: it names the artifact prefix and is stamped on every recommendation, so changing one orphans the artifacts already trained under the old name.
A trained model, identified by the horizon its states were assembled at.
Two per hotel, differing in training window and nothing else — same features, phases, policy class, hyperparameters, guardrails and masking. Further divergence should be earned from the shadow data, not assumed.
The value is a storage key: it names the artifact prefix and is stamped on every recommendation, so changing one orphans the artifacts already trained under the old name.
Possible values: [H90, H365]
"H90"