Skip to content

ADR-v2-121 — LoadGuard (cognitive-load-aware guardrails)

Status: Accepted (2026-07-03) · Wave N Context links: cmdsafety (command risk), reask (confidence re-ask), coach (analytics), voicehealth, [[adr-011]]

Context

Wave N research (#7) — speech carries measurable cognitive-load signals: more/longer pauses, higher filler rate, slowed speaking rate, more self-corrections. When a user is overloaded (stress, ADHD, fatigue), that is exactly when a mis-fired destructive voice command hurts most. The set gates what is risky (cmdsafety) and how confident the ASR is (reask), but not the user's state. Anchor: cognitive-load-from-speech literature (arXiv 2606.12971); ADHD "Understood" guidance.

Decision

Add an opt-in LoadGuard: [loadguard] enabled=false, threshold=0.7. Pure cores in loadguard/policy.py: estimate_load(metrics) — normalize whichever signals the pipeline already computes (pause_ratio, filler_rate saturating at 0.15/word, slowdown below 160 wpm, self_corrections saturating at 3) to 0..1 and average them (missing signals skipped); guard_policy(load, threshold) — high (≥ threshold): widen confirmations + defer risky actions; elevated (≥ 0.7·threshold): widen confirmations only; else normal. OFF by default.

Consequences

  • Destructive actions get harder to trigger exactly when the user is least reliable.
  • Pure estimator + policy → fully testable; signals reuse existing pipeline metrics (no new audio processing, no new dependency).
  • Distinct from cmdsafety (action risk) and reask (ASR confidence) — this models user state.
  • Privacy (ADR-011): scalar features only, computed locally, never stored raw.
  • Caveat: heuristic signal fusion (not a trained model); per-user baselines are a later tier; off by default.