Skip to content
CLEVYA CLEVYA Docs

Egress log

The proof of what went out

Every time a request goes to a language model, CLEVYA records a trace of it in an egress log. This log answers a simple question that a data protection officer (DPO) or an auditor will ask: what actually went out of our server, and in what form?

What the log contains

  • The fingerprint (hash) of what was sent, not the real data.
  • The agent, the workspace and the timestamp.
  • The reason for any blocking decision (a source forbidden from the cloud is never sent, even if it would be relevant: permission takes precedence over relevance).

What an entry looks like

TimestampTargetAllowed sourcesMasked identifiersFingerprint (hash)Status
2026-06-25 09:14:02Cloud (Anthropic)src#A12, src#A1979f2c…b41esent
2026-06-25 09:15:37Local (Ollama)src#A0731a0d…77c2sent
2026-06-25 09:16:08Cloud (Anthropic)src#A33—c4e8…02afblocked (source forbidden from cloud)

The “Fingerprint” column is the hash of what was sent, not the content. “Masked identifiers” is a plain counter. A blocked row sends nothing: it records the reason (“source forbidden from cloud”), because permission takes precedence over relevance.

Why it is verifiable

The entries are chained: each entry references the fingerprint of the previous one. An entry modified or deleted after the fact breaks the chain, which makes tampering detectable. The root of the chain can be signed. Verification can be recomputed on the browser side, without trusting the server.

This is what makes the log enforceable: it serves as evidence in a data protection impact assessment (DPIA) or in a record of processing activities (Article 30, GDPR record of processing).

Honesty

The log proves what went out, not that nothing else could ever go out through a non-instrumented path. The guarantee rests on the fact that every path to a model passes through the same egress control before sending.

Going further