Data lineage

For every agent run, Trefur records a content-addressed fingerprint of the input and output, the model version, the prompt template version, and the tool versions that ran. Drift is detected against this canonical baseline.

What we hash

• Input schema hash — SHA-256 of the canonical JSON form of the agent input. Hashes the shape, not the bytes — so a different customer name with the same field set produces the same input-schema hash.
• Output schema hash — same approach on the response.
• Prompt fingerprint — SHA-256 of the canonical-JSON form of the message array. Gated by your Prompt capture setting (off / hash-only / redacted / full). Hash-only is the default.
• Tool versions — for every tool call, the declared tool version that ran.

What drift looks like

Drift is the divergence between an agent's current rolling signature and its recent rolling baseline. Trefur emits a drift event when the signature shifts beyond the configured threshold:

• New input fields appear.
• Output schema collapses (an optional field disappears).
• The model is silently upgraded by the provider.
• A tool's declared interface changes.

Privacy

We do not store the raw input + output bytes unless your tenant's Prompt capture setting is set to full. Default is hash-only — the hashes give you drift detection without retaining sensitive content.