Autonomy
Trust is earned with numbers, not feelings. Breach a gate — drop a level, automatically.
Level 2 — Staged autonomy
unlocked Oct 824h staging buffer; silence approves dispatch; exceptions to humans.
Consecutive clean approvals143 / ≥ 100
Bounce rate0.8% / < 2%
Spam complaints0 / 0
Qualified replies12 / ≥ 3
Reply-classification audit (50-sample)98% accurate / stable
The ladder
L0
Supervised
Every draft individually approved before send.
L1
Batch-supervised
Daily batches one-click approve; outliers held individually.
L2
Staged autonomy
24h staging buffer; silence approves dispatch; exceptions to humans.
L3
Narrow auto-send
Pre-approved templates, strict ICP slice, daily cap. Holdout-validated.
Circuit breakers
Bounce > 3% / mailbox / 48hauto-pause + demote
Any spam complaintauto-pause + alert
FR-bound or blocked-country send attemptinstant global halt
2 consecutive negative repliespause sequence
Evidence-quality regression (QA sample)demote one level