"HITL Doubt Loop (SSOT)"

HITL Doubt Loop (SSOT)

This is the Single Source of Truth (SSOT) for the Human-In-The-Loop (HITL) Doubt Loop architecture. It defines how autonomous agents express uncertainty, how humans intervene, and how safe skepticism is rewarded.

1. Triggering Doubt

Agents request human intervention via the vox_doubt_task MCP tool.

  • This immediately transitions the task state to TaskStatus::Doubted.
  • The system fires a TaskDoubted event to the vox-orchestrator event bus.

2. The Resolution Agent

When a TaskDoubted event is detected, the ResolutionAgent (living in the vox-dei crate) takes control.

  • It pauses all automated execution streams for the affected task.
  • It engages the FreeAiClient to assist the human in resolving the ambiguity.
  • It tracks the resolution budget via BudgetManager.

3. Audit Report Format

Upon resolution, the ResolutionAgent must submit an audit report.

  • The report logs the nature of the doubt, the human's input, and the cost incurred.
  • It differentiates between "legitimate ambiguity" and "AI obsequiousness".

4. Gamification Hook (vox-ludus)

The audit report is sent to the vox-ludus gamification crate.

  • If the doubt was raised due to detected obsequiousness or true capability gaps (healthy skepticism), the internal_affairs achievement trigger is fired.
  • The agent earns xp for avoiding hallucination.

5. LML Escalation Path

The HITL doubt loop is also the terminal escalation state when the proposed LLM Mediation Layer (LML) exhausts its repair-loop budget. When RepairPolicy.max_attempts is reached without a valid validated output, the LML calls vox_doubt_task on behalf of the current task.

See research-llm-output-mediation-validation-2026.md §6.3 and §11 (Wave 1) for the design of the repair loop and escalation trigger.