confide
Local-first de-identification toolkit — eight tools that redact, verify, restore, and store therapy transcripts without anything leaving your machine.
CONFIDE is a local-first de-identification toolkit packaged as a Claude Code plugin with eight namespaced skills (confide:setup, anon, red, rehydrate, view, audit, vault, annotate). It redacts PII from session transcripts with a layered stack (regex → Natasha RU NER → local LLM), emits a reversible map kept only on-device, checks residual re-identification risk, restores real values into a cloud analysis locally, and stores raw data behind three locks. Built from the bilingual RU/EN CONFIDE benchmark.
What it does
CONFIDE turns the bilingual RU/EN CONFIDE benchmark into a working toolkit: eight namespaced skills that de-identify session transcripts (therapy, coaching, supervision, consulting) and keep the secrets local. The detection stack runs in layers — regex for emails/URLs/phones/IDs/dates, Natasha for Russian named entities, and a local LLM (Ollama qwen2.5:3b) for quasi-identifiers — merging their spans into a single redacted GREEN copy. Nothing is sent to the cloud unless you explicitly opt in on already-redacted text.
The toolkit’s defining move is reversible redaction: confide:anon writes a structured map (<name>.map.json) that lives only on your disk, gitignored, never shipped. Redact a transcript, run the safe GREEN text through any cloud model, then confide:rehydrate puts the real names back locally — completing the round-trip without the cloud ever seeing a real identifier.
The eight tools
- confide:setup — install the dependency stack, pull the local model, write optimal defaults so the rest works with zero config.
- confide:anon — redact PII locally and emit a reversible map (reserved sentinel grammar
[CONFIDE_PERSON_0001]), with a counts-only stats summary. - confide:red — residual re-identification risk check on redacted output: what an attacker could still single out, infer, or link (GDPR Art-29), qualitatively.
- confide:rehydrate — restore real values into a cloud analysis locally; robust to LLM placeholder mangling, never touches ordinary prose, idempotent.
- confide:view — a self-contained interactive HTML diff of original ↔ redacted ↔ restored, colour-coded by type, with All / None / Selected toggles.
- confide:audit — stats-only PII scan across a folder of sessions → aggregate report; never prints a transcript or a real value.
- confide:vault — operationalize the three locks (FileVault + encrypted store +
sops/ageper-file) for raw data, with a status checklist.
- confide:annotate — build a PII gold set with human annotators: a zero-install browser tool, a codebook, inter-annotator agreement (κ), and adjudication.
Key features
- Local-first by default — raw text and the reversible map never leave the machine. Cloud use is opt-in and only on already-redacted GREEN text.
- Reversible round-trip — redact → cloud-analyze the safe text → rehydrate locally. The map is the only artifact with original values, and it stays on disk (0600, gitignored).
- Russian + English — Natasha is native Russian NER; the regex and LLM layers handle both, including morphological variants.
- Honest risk framing —
confide:redsurfaces residual risk categories, not a recall score, with the caveat that absence of a finding ≠ safety. Dual-use guardrails: own data only. - Stats-only at scale —
confide:auditreports counts and distributions across a corpus without ever emitting a name. - Reproducible — engine-agnostic transport (Ollama default, llama.cpp optional); 105 offline-deterministic tests; round-trip and trigger evals.
When to use
Before sending any session transcript to a cloud model — pipe it through confide:anon, send the GREEN copy, then confide:rehydrate the result. Use confide:red to sanity-check what’s still inferable, confide:audit to measure PII across a whole corpus, confide:vault to store raw data behind three locks, and confide:annotate to build a labelled gold set for evaluating any de-identifier. Useful for supervision prep, research datasets, and compliance with 152-FZ, GDPR, or HIPAA.
Built from the CONFIDE benchmark. This is the tool facet — run on your own data — not the scored benchmark. The plugin ships synthetic data only.