💡 Contestability Engine: Avatar-Mediated AI Decision Challenge Under EU AI Act Article 14

[github-actions[bot]]

Summary

Enable users to challenge any agent decision conversationally through the avatar. When a user says "I disagree" or "why did you do that?", the avatar explains its reasoning from the agent's hidden thinking, presents alternative approaches, and logs the contestation with full audit trail. This implements EU AI Act Article 14 (human oversight) and makes TalkTerm the first desktop agent with formal contestability support.

Market Signal

EU AI Act full enforcement began August 2, 2026 — penalties up to €35M / 7% global turnover. Article 14 requires human oversight including the ability to contest AI decisions. The Vector Institute released an open-source Agentic Transparency framework for auditable agent trajectories. MiniOrange and Seekr research confirms audit trails are "make-or-break" for enterprise AI adoption. No desktop agent currently implements formal contestability.

User Signal

• Enterprises adopting AI agents need governance and compliance documentation (79% face adoption challenges despite high investment — WRITER 2026 survey)
• Non-technical users need explanations they can understand, not raw logs or technical reasoning
• TalkTerm's existing confirm-plan pattern (FR20) gates destructive actions but does not allow users to challenge COMPLETED decisions or understand agent reasoning post-hoc

Technical Opportunity

Extends existing audit repository (SessionRepository, AuditRepository ports) and confirm-plan pattern (FR20). Agent thinking is already captured but hidden per FR22 — contestability selectively surfaces it through avatar narration. Challenge events become a new audit log entry type. The AgentBackend abstraction layer can expose decision metadata without coupling to the Claude SDK.

Key extension points:

• ContestabilityEvent domain type in src/shared/types/domain/
• Extended AuditRepository port with challenge event recording
• IPC channel agent:contest for challenge flow
• Avatar narration of reasoning via the existing overlay/caption system

Assessment

Dimension	Score	Rationale
Feasibility	med	Requires extending audit pipeline and exposing agent thinking selectively; rollback limited to reversible actions
Impact	high	EU AI Act compliance, enterprise differentiator, trust-building for non-technical users
Urgency	high	EU AI Act enforcement began August 2, 2026 — compliance window is NOW

Adversarial Review

Strongest objection: Undo for agent actions that modified files or called APIs is extremely difficult. You cannot un-send an email or un-write a file that another process has read.

Rebuttal: Contestability does not require full undo — it requires explanation, alternatives, and audit trail. For reversible actions (file writes, branch creation), TalkTerm can offer rollback. For irreversible ones (API calls, external writes), it provides the reasoning, records the challenge, and redirects the agent on an alternative path. This satisfies the EU AI Act requirement and is more than any competitor offers.

Suggested Next Step

Design a ContestabilityEvent domain type in src/shared/types/domain/ and extend the AuditRepository port to record challenge events. Define how agent thinking metadata flows through the IPC bridge for avatar narration.