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
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_appliedmeans 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.