Human in the Loop

[joshua-temple]

The Proposal

Define explicitly where human judgment is required in automated workflows, and make those checkpoints fast. Automation with oversight, not automation as abandonment.

Why This Matters

Full automation fails in two directions. Too little oversight and irreversible mistakes ship: a bad external email, a destructive migration, a compliance breach. Too much and humans become the bottleneck, review degrades into rubber-stamping, and the value of automating evaporates.

The root failure is rarely the automation itself. It is an undefined boundary. When nobody has written down which decisions need a human, every reviewer draws the line differently, and the line drifts toward whatever is convenient that day. An explicit, agreed checkpoint map makes oversight deliberate instead of accidental.

When to Require a Human

• High stakes: hard-to-reverse actions, external communication, significant spend, safety-critical changes
• Low confidence: high model uncertainty, novel situations, conflicting signals
• Policy: compliance mandates, audit trails, formal approval workflows

Making Review Efficient

• Surface the relevant context alongside the decision; never make the reviewer dig
• Present a clear recommended action, not an open-ended question
• Make approve/reject a single step
• Batch similar low-risk decisions into one review

Open Questions

1. Calibration signals. What measurable indicators tell us a checkpoint is mis-tuned: approval rate near 100% (rubber stamp), queue latency past a threshold (bottleneck)? Who owns re-tuning it?
2. Detecting theater. How do we verify review is real: sampled audits of approved decisions, tracking time-per-review, occasionally injecting known-bad items to confirm they get caught?
3. Escalation with no reviewer. When confidence is low and no human is available, does the system block, fall back to a safe default, or queue? Per checkpoint or one global policy?
4. Where the map lives. Should checkpoint definitions be code (versioned, reviewable, testable) rather than tribal knowledge, and what format makes them auditable?

[joshua-temple]

Accepted. Documented in #13.