🐉

Right to Explanation & Redress

EU AI Act Art 86 · GDPR Art 22(3) · Individual Decision-Making Explanation · Contestability · FRIA Safeguards · EAT Directive aligned

EU AI ACT ART 86 GDPR ART 22 ACTIVE

Explanation Tiers

30d

Explanation SLA

Redress Steps

Escalation Levels

100%

Decisions Loggable

Black-Box Outputs

What is Article 86?

Article 86 of the EU AI Act grants affected persons the right to explanation of individual decision-making made by high-risk AI systems. When a high-risk AI system outputs a decision that produces legal effects or similarly significantly affects a person's health, safety, or fundamental rights, that person has the right to obtain:

Clear and meaningful explanation of the decision's output
The main elements of the decision logic — what factors were weighed
The rationale behind the decision as it applies to them specifically
The ability to contest the decision through a defined redress process

This right intersects with GDPR Article 22(3) (right to obtain human intervention, express one's point of view, and contest the decision) and Article 15 (right of access, including the logic of automated processing). DEFONEOS implements this through a structured, multi-tier explanation framework backed by the SIGIL chain's immutable audit trail.

5-Tier Explanation Framework

Tier 1 — Plain Language Summary

The decision is expressed in plain English at a reading age of 12. No technical terms. No acronyms. Example: "Your application for housing priority was ranked below 14 other households because your current accommodation meets the statutory minimum standard and your household size places you in Band C."

Tier 2 — Decision Factors

The specific factors that influenced the decision are listed with their individual contribution weights. Each factor is traceable to the source data and the governance rule that authorised its use. Example: "Factor: Household size (weight: 0.23). Source: Housing register entry dated 12 Jun 2026. Governance: Housing Allocation Policy §4.2."

Tier 3 — Decision Logic Trace

A step-by-step trace showing how the factors combined to produce the output. This includes the model's reasoning path, any human override points, and the BFT council vote if one was held. Each step carries a SIGIL hash for tamper-evidence.

Tier 4 — Counterfactual Explanation

What would have needed to be different for the decision to go the other way? This is generated automatically from the model's decision boundary analysis. Example: "If your household included a dependent under 16 with a registered medical need (Band A factor), your application would have been ranked in the top 5."

Tier 5 — Full Technical Audit Trail

Complete SIGIL chain entry with Ed25519 signature, model version, training data provenance, inference timestamp, all factor weights, confidence score, calibration metric, BFT vote tally, and human review logs. Available to regulators, independent auditors, and the data subject's legal representative on request.

7-Step Redress Process

Step	Action	SLA	Responsible
1	Request — Affected person submits explanation/redress request through the DEFONEOS portal, in writing or verbally (phone/in-person, transcribed by agent)	Immediate	Defoneos Agent / DPO
2	Acknowledge — System confirms receipt, assigns case ID, and provides the Tier 1-2 explanation automatically within 72 hours	72h	Automated
3	Review — A qualified human reviewer examines the decision, the model output, and the affected person's submissions. The reviewer must NOT be the person who made or approved the original decision.	14d	Human Reviewer (SC-cleared)
4	Explanation — Full Tier 1-4 explanation delivered to the affected person, including counterfactual. Technical Tier 5 available on request.	21d	DPO / Reviewer
5	Remedy Options — If the decision is found to be flawed, biased, or not properly explainable, the reviewer can: (a) reverse the decision, (b) request a new decision with corrected data, (c) escalate to BFT council, (d) refer to human appeal tribunal.	21d	Reviewer + DPO
6	BFT Escalation — For decisions with significant legal effects, a 12-of-33 quorum BFT council reviews the case. The council includes the independent reviewer agent (Aegis), the DPO agent, and the care-validation agent. Red line violations are auto-flagged.	30d	BFT Council (quorum 12/33)
7	External Referral — If internal redress is exhausted, the affected person is informed of their right to: ICO complaint (GDPR), judicial review, Parliamentary and Health Service Ombudsman, or the EU AI Office (for AI Act breaches). All evidence from the SIGIL chain is exportable.	As needed	DPO

3 Escalation Levels

Level	Trigger	Response	Timeline
L1 — Standard	Affected person requests explanation within 30 days of decision	Tier 1-2 auto-generated, Tier 3-4 within 14 days	14d
L2 — Expedited	Decision affects housing, benefits, liberty, child welfare, or medical triage	All tiers within 72 hours, human reviewer assigned immediately	72h
L3 — Critical	Potential red line violation, safeguarding concern, or ICO complaint filed	BFT emergency council (quorum 12/33), DPO personally notified, decision suspended pending review	24h

EU AI Act Art 86 Compliance Matrix

Requirement	Status	How DEFONEOS Implements
Right to obtain clear and meaningful explanation	✅ MET	Tier 1-2 plain language auto-generated from decision factors
Main elements of the decision logic	✅ MET	Tier 2-3 factor breakdown with governance provenance
Rationale for the specific decision	✅ MET	Tier 4 counterfactual + Tier 3 logic trace
Right to contest / redress	⚠️ PARTIAL	7-step process designed. Not yet tested with live data subjects. External referral pathway documented.
Human review of decisions	✅ MET	Step 3 requires independent SC-cleared human reviewer
Audit trail for regulators	✅ MET	Tier 5 SIGIL chain with Ed25519 signatures, exportable to ICO/EU AI Office
Timely response	✅ MET	72h SLA for L2 expedited, 14d standard, 30d maximum
Accessible to vulnerable persons	⚠️ PARTIAL	Plain language tier designed. Layperson UX testing not yet conducted. No non-English language support yet.

Counterfactual Generation

def generate_counterfactual(decision_id: str, model_version: str) -> Counterfactual:
    """Generate what-would-need-to-change for decision to flip."""
    decision = sigil_chain.get(decision_id)
    factors = decision.factors  # {factor_name: weight}

    # Sort by absolute weight descending
    sorted_factors = sorted(factors.items(), key=lambda x: abs(x[1]), reverse=True)

    counterfactuals = []
    for factor, weight in sorted_factors[:5]:  # Top 5
        # Compute decision boundary distance
        delta = compute_decision_boundary(model_version, factor, decision.threshold)
        counterfactuals.append(Counterfactual(
            factor=factor,
            current_value=decision.input_values[factor],
            required_value=delta.flip_point,
            governance_rule=find_governance_rule(factor),
            plain_language=translate_to_plain_english(factor, delta.flip_point),
        ))

    return Counterfactual(
        decision_id=decision_id,
        explanation="If any of the following were true, the decision would have been different:",
        changes=counterfactuals,
        sigil_hash=sign_ed25519(counterfactuals),
    )

Cross-Framework Alignment

Framework	Article	Requirement	DEFONEOS Mapping
EU AI Act	Art 86	Right to explanation of individual decision	5-tier explanation framework
GDPR	Art 22(3)	Human intervention, express point of view, contest	7-step redress process steps 3-7
GDPR	Art 15(1)(h)	Right of access to logic of automated processing	Tier 3 logic trace + Tier 5 full audit
GDPR	Art 13(2)(f)	Information about automated decision-making	Pre-decision privacy notice
ECHR	Art 6	Right to a fair trial / due process	Step 7 external referral to tribunal
ECHR	Art 8	Right to respect for private and family life	Sovereign storage, no foreign transfers
UK DPA 2018	Schedule 1 Part 2 §10	Safeguards for automated decision-making	Full Art 86 + GDPR Art 22 implementation
HRA 1998	S.3	Interpretation compatible with Convention rights	External referral pathway to courts
ICO Guidance	Automated DM	Meaningful information about the logic	Tier 2-4 explanations with plain language
EU AI Act	Art 27 FRIA	Fundamental rights impact assessment	Explanation framework is FRIA output

Integration with DEFONEOS Architecture

Every DEFONEOS decision is logged to the SIGIL chain with the following fields, enabling instant explanation generation:

decision_id — Unique Ed25519-signed hash
model_version — Exact model checkpoint and training data provenance
factors — Dict of {factor_name: {value, weight, source, governance_rule}}
threshold — Decision boundary threshold
confidence — Calibration score (Brier score ≤ 0.15)
bft_vote — Council vote tally (if triggered) with per-voter rationale
human_reviewer — Identity of reviewer (if triggered) and decision
red_line_check — 7 red lines checked, all must be CLEAR
sigil_hash — Hash-chained to previous entry, tamper-evident

⚠️ Honesty Register

The explanation framework described here is the designed system. DEFONEOS has not been deployed in a live public-service context where affected persons have exercised their Article 86 right through this system. The 72h/14d/30d SLAs are design targets, not measured performance. The counterfactual generation algorithm is implemented as a prototype but has not been validated against the EU AI Office's forthcoming guidance on "meaningful explanations." The BFT council escalation (Step 6) is functional in the sovereign substrate but has not been tested with a live redress case. The "plain language at reading age 12" claim is a design specification, not a tested outcome — layperson UX testing is a documented human gate. No DPO has been formally designated for DEFONEOS under GDPR Art 37.