Activation inspection

Inspection records selected internal activations while the model processes specified prompts. It does not optimize weights, create a trained adapter or prove that a neuron causes a behavior. It helps choose coordinates to examine with controlled interventions.

For example, run the same short product question with and without a product name and inspect selected neurons over the actual prompt tokens. Compare the same checkpoint, tokenizer and token positions.

Set up the experiment

Choose New → Inspect activations in Experiments.

The current editor accepts up to eight prompts and 64 coordinate entries. Coordinates and token positions belong to the exact model revision; indices from a different checkpoint are not a feature dictionary. Capture and forward work limits prevent an apparently small request from returning an unbounded activation tensor. Increasing generated tokens also increases work.

Read and use the result

Inspect token-aligned values, capture limits and actual runtime diagnostics. A large activation is an observation, not an automatic semantic label. Check several prompts and a baseline before naming a neuron or assigning a role. Missing or incomplete capture is not a zero activation.

From a completed compatible hosted inspection, select actual Off-observed neurons and choose Train selected neurons. The existing neuron-training editor receives the exact model, result and selected coordinates. It still requires reviewed target/background data and separate training approval. Imported observations do not grant hosted training authority.

To test an inference-time intervention instead, use Neuron steering. Inspection itself has no weight artifact to save or serving capacity to activate.

Review, run and recover

Use Review → Prepare review to resolve the exact inputs, compatible runtime and maximum spend. Inspect the retained preparation before approving Start. Opening the editor, changing tabs and saving a draft do not start paid compute. Preparing a run is separate from starting it.

If a response is lost, reopen the same experiment and retained execution before retrying. Preserve its operation identity and revision; creating a second run can duplicate work and charges. Use the existing refresh, retry or cancellation controls. Cancellation is a request: follow the execution through cleanup and inspect remaining capacity and funding before treating it as finished.

Runtime discovery is authoritative. A selectable action or saved configuration does not prove that a checkpoint/image/GPU combination is qualified. Unsupported combinations cannot start simply because sufficient credits are available.

API, SDK and harness

Use compileModificationExecution, prepareModificationExecution and startModificationExecution with the inspection operation. CLI equivalents are training hosted-compile-modification, hosted-prepare-modification and hosted-start-modification; inspect the retained modification execution and comparison operations afterward.

Use the official CLI with your configured account and workspace. Run openpond training --help and the selected command's help for the current request schema; start from the SDK's typed draft/request builders rather than inventing an unversioned JSON payload. Web, API, TypeScript/Python SDKs and the CLI retain the same preparation, execution and Model identities. Paid Start and serving activation require explicit approval; a harness should retain these identities for recovery.