Identity Guard · Privacy Inspector

Sensitive values are tokenized in your browser before they reach the model — restored only locally.
protected token (what the model sees) restored value (only on your device)
◈ POC status Checkpoints A·B done · C/D/E partial · 136 tests green · 4 PII types · collapse me for the demo →
Progress
  • A — redaction engine done
  • B — mock POC + inspector done
  • C — real-stream (gpt-5.5) partial
  • D — evals / perf / privacy D1·D3·D4a
  • E — contextual model spike design
  • Tests 136 green
PII filtered now
  • Email on
  • US SSN (validated) on
  • API keys / secrets (9 prefixes) on
  • Payment card (Luhn) flagged
  • Phone · URL · names · addresses deferred
Detector evals (synthetic corpus)
  • Precision / Recall / F1 1.00
  • False positives / 1k 0
  • Exact span boundary 1.00
  • Categories scored EMAIL·SSN·API_KEY·CARD
Model protocol (gpt-5.5)
  • Token preservation 100%
  • Hidden-value leakage 0%
  • Unnecessary refusals 0%
  • Task success vs off +38%
Performance
  • protect() p99 @ 50 KB ~0.42 ms
  • Catastrophic backtracking none
  • Redaction core (gzip) ~4.4 KB
  • Runtime dependencies 0
Guarantees
  • Fail-closed on detector error ✓
  • Exact vault-keyed restore only ✓
  • Vault in-memory, never serialized ✓
  • Multi-turn re-protection ✓
ON ▍ streaming… Model
Real-world examples — the model helps, but never sees the secret
Adversarial / limit cases — the model can’t see the value, so it should refuse, not guess

Privacy Inspector — per turn: what crosses the wire, and what came back