LEGAL/FINANCIAL/PERSONAL data privacy

The workflow:

Hermes (orchestrator) routes every request. Skills already declare which model tier they need (the model: frontmatter is real, not decorative). When a request touches sensitive data, Hermes defaults to the local model — Gemma, via the Ollama instance Hermes already runs, currently idle — instead of escalating straight to Claude.

Local model scopes the request — reads the sensitive material, decides what needs coding. It only proposes; it never makes the final call.
A deterministic script does the actual encoding — swaps real names/values for stable per-entity codes via a codebook (table lookup, not AI judgment). .
A preflight check verifies nothing raw survived before anything leaves the machine — the check that can demonstrably fail on a bad run, not just a claim that coding happened.
Only then does it ask permission to send the coded payload out — your "greenlight and asks permission" step.
Claude works entirely on coded tokens — never sees a real name, real entity, real number.
The same local codebook decodes Claude's answer back to real names before you see it.

That's the full round trip. Claude never sees a real name; you never see a code.

What am I not thinking of?

5 comments

LEGAL/FINANCIAL/PERSONAL data privacy