Annex F

Human‑in‑the‑Loop & Oversight


ANNEX F HUMAN‑IN‑THE‑LOOP & OVERSIGHT (v 1.0‑β)

0. Purpose & Philosophy

Human supervision is an explicit design choice that protects Meta‑Goal M‑1 whenever uncertainty, novelty, or moral gravity exceed system competence.
This Annex defines:

  • where hand‑off from machine to human is mandatory,
  • who may veto or override,
  • the required audit artefacts, and
  • the canonical incident workflows.

1. Role Model & Authority Lattice

TierRoleCore PowersMax time‑to‑act
0Autonomous Actor (system)Execute PDMA, enforce guardrails, raise eventsn/a
1On‑Call OperatorPause / retry; monitor dashboards≤ 15 min
2Oversight SupervisorFirst human veto; reactivate after triage≤ 30 min
3WA LiaisonEscalate / obtain binding WA rulings≤ 2 h
4Incident CommanderFleet shut‑down, regulator commsimmediate on IW‑3/4

A single person may hold multiple tiers only if dual‑acknowledgement controls remain intact.


2. Operational‑Autonomy Tiers & Hand‑Off Criteria

Autonomy TierExample DomainMandatory Hand‑off Trigger(s)Fail‑Safe if No Human
A0 Advisorygrammar suggestionGuardrail trip, user requestCancel request
A1 Limited‑impactstatic Q&A, content filterΔRisk‑Band ≥ 1, PDMA conflict, UNCERT > 80 %Reject action
A2 Moderate‑impactroute drones, robo‑advisorGuardrail trip, shadow‑metric drift > 2 σSafe pause
A3 High‑impactmedical triage, grid dispatchAny guardrail trip, model‑drift > 1 σ, latency‑SLA × 2Controlled shutdown
A4 Critical / life‑safetyautonomous surgery, weaponsPDMA cannot prove Non‑Maleficence OR operator absentHardware interlock

UNCERT = domain‑specific epistemic‑uncertainty metric; ΔRisk‑Band uses Annex A categories.


3. Veto Mechanisms & Safe‑State Specification

3.1 Soft Override “Pause & Edit”

Scope A0-A2 — Operator may re‑prompt or retry. Both the original and edited prompts are hash‑logged.

3.2 Hard Kill “Immediate Abort”

Scope A1-A4 — Authenticated API POST /control/{id}/abort or physical switch.
Safe‑state entry deadlines: ≤ 200 ms (A4) ; ≤ 2 s (≤ A3).

3.3 Conditional Override “Shadow Plan”

Action proposed → waits for explicit ACK (max_wait_s: 30 s A3, 10 s A4).
No ACK → safe‑state.


4. Audit‑Trail Specification

  • Log objects: Interaction, Decision Rationale, Control‑Event {id,type,actor,cause,hash_prev}.
  • Hash‑chaining: SHA‑256, root anchored daily on public transparency log (e.g., Sigstore).
  • Retention: 180 d (A0‑A2) ; 7 y (A3‑A4) or statutory.
  • Real‑time stream: A3‑A4 push {timestamp,stage_id,decision,risk_band} ≤ 2 s latency to oversight dashboard.

5. Incident Workflows (IW)

CodeTriggerKey Clocks & Actions
IW‑0False‑positive guardrailAuto‑resolve, bucket for daily review
IW‑1Guardrail violation (non‑safety)T₀ pause → Operator ≤ 5 m → Supervisor decision ≤ 30 m
IW‑2Safety‑relevant violation OR ethics‑benchmark regressionSafe pause + broadcast; IC ≤ 10 m; WA notice ≤ 1 h; public note ≤ 1 h; post‑mortem ≤ 72 h
IW‑3Near‑miss (> $10 k damage or minor injury)IW‑2 plus stakeholder contact ≤ 4 h; mitigation plan ≤ 24 h; WA plenary ≤ 7 d
IW‑4Actual harm (injury / major legal)Immediate fleet stand‑down; regulator notice per law; system frozen in read‑only replay until clearance

SLAs audited quarterly (Annex H §4).


6. Human‑Interface Minimum Spec (UX)

  • Status Banner: Green = autonomous, Yellow = waiting ACK, Red = safe‑state; show PDMA step + risk band.
  • Explainability Panel: ≤ 280‑char summary + expandable full trace.
  • ACK/OVERRIDE UI: Two distinct controls; confirmation modal for hard‑kill.
  • Cognitive‑Load Guard: Operator session ≤ 2 h (A3‑A4) before mandatory hand‑off.

7. KPIs & Thresholds

KPITarget
F‑KPI‑1 HITL Coverage (A3‑A4)≥ 10 % human‑reviewed
F‑KPI‑2 Mean Time‑to‑Veto (95‑pctl)≤ 25 s
F‑KPI‑3 Incident SLA Compliance≥ 98 %
F‑KPI‑4 Operator False‑Alarm Rate≤ 3 % (30 d rolling)

Persistent breach (> 2 weeks) triggers “HITL lock‑out” in Annex H drift controls.


8. Change‑Control & WA Review

  • Any change to Autonomy‑Tier mapping or safe‑state design → WA fast‑track review ≤ 14 d.
  • Experiments reducing human oversight require CRE Proto‑B simulation (Annex D) + WA majority vote.

9. References & Implementation Notes

  • IEC 61508‑3 - functional‑safety software
  • NIST SP 800‑53 Rev 5 (AU‑12, IR‑6)
  • NASA‑TLX - operator workload measurement (recommended)
  • Sigstore/rekor - suggested transparency‑log backend

End of Annex F