Method Authoring

NeuroMouse methods are small plugins that run against a validated Dataset. A method declares the contract fields it consumes, the output fields it promises to return, and an optional panel hint for rendering.

Use these two files first:

• full working example: band_power_summary.py
• starter template: method_template.py

Author Surface

The SDK method surface is intentionally small:

class MyMethod:
    name = "my_method"
    params_type = MyMethodParams
    required_inputs = ("meta.channels", "welch_psd.frequencies", "welch_psd.psd")
    output = OutputSpec(...)

    def compute(self, dataset: Dataset, params: MyMethodParams) -> dict[str, Any]:
        ...

The current primitives live in packages/sdk/src/neuromouse_sdk/method.py:

• OutputField(path, description="", unit=None)
• PanelSpec(id, title, kind, field)
• OutputSpec(fields, panel=None)
• Method protocol
• build_params() for dataclass or Pydantic-style params

Registry Contract

Methods run through MethodRegistry, not by direct viewer calls. The registry:

• validates the method declaration
• verifies every required_inputs field path exists on the dataset
• coerces params into params_type
• calls compute()
• verifies every declared output.fields path exists in the returned mapping
• verifies output.panel.field points at one declared output field

That means a method cannot silently depend on missing data and cannot return a shape the UI has to guess.

Minimal Workflow

1. Copy method_template.py.
2. Rename the class, name, and output root key.
3. Define params as a frozen dataclass or Pydantic-compatible type.
4. Declare the exact contract fields in required_inputs.
5. Declare every returned field in OutputSpec.
6. Add a panel only when the result has a clear table, chart, or summary surface.
7. Register and run through neuromouse_core.method_registry.MethodRegistry.
8. Add tests that cover declaration shape, params coercion, output fields, and at least one realistic dataset.

Example Reference

band_power_summary integrates PSD values across a frequency band. It demonstrates the expected shape:

• params: BandPowerParams(min_hz=8.0, max_hz=13.0)
• inputs: meta.channels, welch_psd.frequencies, welch_psd.psd
• outputs: band metadata, per-channel rows, mean power, and top channel
• panel: table over band_power_summary.channels

The tests in packages/sdk/tests/test_method_author_api.py and packages/core/tests/test_method_registry.py are the best executable examples of the authoring contract.

Review Checklist

• Does the method read only declared input paths?
• Does compute() return every declared output field?
• Does the output root key match the method name or a clear method namespace?
• Are params typed and deterministic?
• Does the method fail with a clear ValueError for invalid scientific settings?
• Is provenance or unit information included when the output would otherwise be ambiguous?