Embedding Noise (Experimental)

Synapse Layer can add Gaussian noise to a derived embedding for experimentation. This is not a formal differential-privacy guarantee, and the formal claim is suspended.

Warning

Claim suspended. The persisted semantic embedding is not yet covered by a formal differential-privacy mechanism. Do not treat privacy_applied as proof of differential privacy.

Status

The formal differential-privacy claim is suspended in public copy, docs, Trust Page, SDK and sales materials until the persisted vector is protected and independently reviewed. The current code path is an experiment; it is not a privacy guarantee.

The next decision is P1.1: implement and validate a reviewed mechanism for the persisted vector, or remove the feature and the claim.

What the experiment does

For a derived embedding used for sizing and audit, the pipeline can add Gaussian noise, normalize the result and record sigma/SNR metadata. This is a noise experiment, not a formal epsilon-delta guarantee for the semantic embedding that is persisted.

The current implementation is useful for testing and instrumentation. It is not evidence of a formal guarantee for the persisted vector.

What we do not claim

  • No formal epsilon-delta differential-privacy guarantee.
  • No claim that the persisted semantic embedding is protected by this noise.
  • No claim that embedding inversion is prevented.
  • No claim that privacy_applied means formal differential privacy was applied.

Boundaries

  • PII is handled separately by the Semantic Privacy Guard.
  • Metadata fields (agent, timestamp, intent) are not protected by this experiment.
  • Any future formal claim requires a reviewed mechanism, tests on the persisted vector, and explicit approval before publication.