Skip to content

Privacy

Jump to bottom
Chuyue Wang edited this page Apr 22, 2026 · 2 revisions

Privacy

Data Boundary

Cortex keeps biometric processing local by design.

  • Webcam frames, landmarks, pulse/HRV, blink, posture, and SQI are processed in-memory.
  • Biometric signals are not sent to LLM providers.
  • Intervention LLM calls only include workspace context (tabs, editor/terminal context, complexity/state labels).

What Is Stored Locally

  • Baselines and calibration artifacts (storage/baselines/)
  • Session and evaluation artifacts (storage/sessions/, storage/policy_log/, storage/reports/)
  • Optional learning metadata (tab relevance, helpfulness, consent state)

Redis is optional; in-memory fallback is available.

Prompt/LLM Safety Guarantees

0.2.0 adds prompt hardening:

  • User and learned strings are sanitized before prompt interpolation.
  • Control characters and suspicious prompt-injection markers are stripped.
  • Lengths are bounded and braces escaped to prevent template breakouts.
  • LLM plans are schema-validated and unsafe actions are degraded/dropped before execution.

Microrandomization / AMIP Policy Logging

When eval.policy=amip, Cortex logs decision tuples for causal evaluation:

  • context features (numeric)
  • action propensity distribution
  • chosen arm
  • linked reward

These records are written locally to storage/policy_log/YYYY-MM-DD.jsonl and summarized in local causal reports.

If you do not want exploration-style policy learning, set:

CORTEX_EVAL__POLICY=greedy

(Or uniform for research-style randomized mode.)

Consent and Autonomy

All impactful workspace mutations are consent-gated. Consent escalation is recency-aware and reversible; high-impact actions (including LeetCode lockout-style interventions) are not executed autonomously without earned trust.

Clone this wiki locally