Use a Manifest

An exported policy package contains a manifest.json file.

exports/act_policy/
├── manifest.json
├── model.xml
└── stats.safetensors

In most cases, you load the package through InferenceModel.

model = InferenceModel("./exports/act_policy")

If you need to inspect metadata directly, load the manifest itself.

from physicalai.inference.manifest import Manifest

manifest = Manifest.load("./exports/act_policy/manifest.json")
print(manifest.model.runner)
print(manifest.model.artifacts)

Declare inference callbacks

Callbacks use the same ComponentSpec syntax and declaration order as preprocessors and postprocessors. Built-in callbacks can use their registered short name with flat constructor arguments, or a fully qualified class path with init_args:

model:
  callbacks:
    - type: rldx1_vtc
      video_length: 4
      video_stride: 2
    - class_path: physicalai.inference.callbacks.LatencyMonitor
      init_args:
        window_size: 100

The RLDX-1 callback above creates temporal image windows before preprocessing without Python-side callback wiring. Manifest callbacks receive on_load, on_predict_start, on_predict_end, and on_reset lifecycle events.

When constructing InferenceModel, an explicitly supplied callbacks list takes precedence over the manifest, including an explicitly supplied empty list. Invalid, unknown, or incompatible callback specifications fail while the model is loading.

Use manifests to describe exported artifacts. Use workflow config to author training, inference, or runtime workflows before execution.