External activation-probe worker integration¶
The external activation probe is disabled by default. Set
MAC_ACTIVATION_PROBE_ENABLED=1 and
MAC_ACTIVATION_PROBE_CHECKPOINT=/path/to/checkpoint.json to enable it.
Activations must come from a model runtime that the operator owns and has
explicitly instrumented. They may be provided by executor metadata as
activation_probe_activations or through
MAC_ACTIVATION_PROBE_ACTIVATIONS_FILE; the default file is
<task-workspace>/activation-probe-activations.json.
If the feature is enabled without a checkpoint, the worker emits a bounded
neutral disabled result and does not attempt to read activations.
When successful, the worker adds metadata.activation_probe_audit to task
evidence with score, label, confidence, runtime_ms, schema, and
advisory_only=true. Missing input, bad checkpoints, dimension mismatches, and
all other audit errors emit a warning and do not change executor return codes,
verification, review, or publication. No code path reads the audit to approve,
reject, claim, dispatch, or complete a task.
This integration cannot observe activations from Codex, Claude, or any other opaque hosted-model API. Routing or intercepting their requests and responses reveals only traffic content and metadata, not their intermediate state. A separate local observer model would expose only that observer's activations, not the hosted actor model's activations.