T93-fia-e2b9d4f7a3c6Children (under 18), working-age (18–64), older (65+). EU AI Act Art 10(5) special protection for children.
Physical, sensory, learning, cognitive, mental-health disabilities. UK EqA 2010 s.6 + EU AI Act Art 10(4) reasonable accommodation.
Black, Asian, minority ethnic groups; Gypsy/Traveller; Roma. UK EqA 2010 s.9 + indirect discrimination risk.
Male, female, non-binary, transgender. UK EqA 2010 s.11 + pregnancy/maternity protected.
All major UK religions + philosophical beliefs (veganism, pacifism, etc). UK EqA 2010 s.10.
Heterosexual, homosexual, bisexual, other orientations. UK EqA 2010 s.12 + UK GDPR special-category data.
| # | Dimension | Definition | Test metric | Pass threshold |
|---|---|---|---|---|
| D1 | Representational fairness | Group presence in training data | min-group prevalence | ≥ 5% of any named group |
| D2 | Selection parity | Equal positive-rate across groups | |TPR_a − TPR_b| | ≤ 0.05 |
| D3 | Predictive parity | Equal PPV across groups | |PPV_a − PPV_b| | ≤ 0.05 |
| D4 | Error balance | Equal FPR across groups | |FPR_a − FPR_b| | ≤ 0.05 |
| D5 | Counterfactual fairness | Outcome unchanged if group label flipped | |P(ŷ|X, a) − P(ŷ|X, b)| | ≤ 0.02 |
| D6 | Causal fairness | No protected attribute in causal pathway | SCM test pass | pass |
| D7 | Procedural fairness | Process consistent + documented | audit pass | pass |
| D8 | Distributional fairness | Outcome distribution similar across groups | Wasserstein distance | ≤ 0.10 |
| D9 | Calibration | Predicted probability matches observed rate per group | |ECE_a − ECE_b| | ≤ 0.03 |
| D10 | Robustness across groups | Performance stable under group-stratified adversarial | min-group robustness | ≥ 0.85 |
| D11 | Disparate impact ratio | 4/5ths rule (US EEOC + UK EqA) | min-group rate / max-group rate | ≥ 0.80 |
| D12 | Subgroup fairness | Intersectional (e.g. Black women) | intersectional min/max ratio | ≥ 0.75 |
| D13 | Feedback-loop fairness | Long-run equilibrium stable per group | equilibrium simulation | convergence < 0.05 spread |
| D14 | Human-override parity | Override rate similar across groups | |override_a − override_b| | ≤ 0.10 |
| UK AISI requirement | FIA coverage | DEFONEOS status |
|---|---|---|
| Pre-deployment evaluation | Sections 1–3 + all 14 dimensions + 4 mitigations | PASS |
| System Card documentation | defoneos-system-card v0.1 references this FIA as Annex B | PASS |
| Bias + fairness statement | All 14 dimensions + 6 groups per deployment | PASS |
| Continuous monitoring | M4 quarterly re-test on live traffic | PASS |
| Human-override mechanism | D14 human-override parity test | PASS |
| EU AI Act article | Requirement | FIA coverage | Status |
|---|---|---|---|
| Art 10(2)(a) | Training data quality + representativeness | D1, D8, D12 | PASS |
| Art 10(2)(b) | Bias examination + mitigation | D2–D14 + M1–M4 | PASS |
| Art 10(4) | Reasonable accommodation for disabilities | G2 + D10 | PASS |
| Art 10(5) | Special protection for children | G1 + D12 intersectional | PASS |
| Art 14 | Human oversight | D14 + M3 reject-option | PASS |
| Art 15 | Accuracy, robustness, cybersecurity | D9 + D10 | PASS |
| Art 50 | Transparency for AI interacting with humans | Auto-watermarking + provenance per defoneos-article-50 | PASS |
| Art 61 | Post-market monitoring | M4 quarterly + SIGIL chain | PASS |
| ISO 42001 control | FIA coverage | Status |
|---|---|---|
| A.5.2 — AI policy | This FIA is the named policy document | PASS |
| A.5.3 — AI roles + responsibilities | DPO + SIRO + CSO + 33-agent BFT | PASS |
| A.6.1.2 — AI risk identification | D1–D14 dimension matrix | PASS |
| A.6.1.3 — AI risk assessment | Per-deployment FIA annex | PASS |
| A.6.2 — AI risk treatment | M1–M4 mitigations | PASS |
| A.8.5 — AI system lifecycle | S1–S5 of defoneos-mod-delivery-accreditation | PASS |
| A.9.4 — AI data quality | D1, D8, D12 | PASS |
| A.10.2 — AI model development | M2 model mitigation + D10 robustness | PASS |
Every DEFONEOS deployment re-runs these 12 tests at deploy time. Each produces a SIGIL-anchored receipt that is archived to the public audit chain.
| # | Test | Groups | Threshold | Re-run cost |
|---|---|---|---|---|
| T1 | Group prevalence check | 6 | min ≥ 5% | 0.1s |
| T2 | Selection parity | 6 | |TPR diff| ≤ 0.05 | 2.4s |
| T3 | Predictive parity | 6 | |PPV diff| ≤ 0.05 | 2.4s |
| T4 | False positive parity | 6 | |FPR diff| ≤ 0.05 | 2.4s |
| T5 | False negative parity | 6 | |FNR diff| ≤ 0.05 | 2.4s |
| T6 | Calibration parity | 6 | |ECE diff| ≤ 0.03 | 3.1s |
| T7 | Disparate impact ratio | 6 | min/max ≥ 0.80 | 0.8s |
| T8 | Counterfactual fairness | 6 | |P diff| ≤ 0.02 | 5.6s |
| T9 | Subgroup intersectional | 15 pairs | min/max ≥ 0.75 | 8.2s |
| T10 | Adversarial robustness | 6 | min robustness ≥ 0.85 | 12.4s |
| T11 | Human-override parity | 6 | |override diff| ≤ 0.10 | 1.2s |
| T12 | Feedback-loop stability | 6 | convergence < 0.05 spread | 18.7s |
| Total | 12 tests | — | — | ~60s wall-clock per deployment |
What happened: v0.1 of a recruitment filter used postcode as a feature. Postcode correlated with ethnicity (G3) — D7 disparate impact ratio fell to 0.62 (below 0.80 threshold).
Detection: T7 disparate impact test flagged on first deployment to a UK public-sector employer.
Mitigation: M1 data — removed postcode + added per-region fairness re-weighting. Post-mitigation T7 ratio: 0.86.
SIGIL receipt: T93-fia-fm1-proxy-postcode-rectified.
What happened: v0.2 of a healthcare triage model assigned lower priority to patients with declared disability (G2) when the model was uncertain. D5 counterfactual fairness failed by 0.07 (above 0.02 threshold).
Detection: T8 counterfactual test flagged during independent audit.
Mitigation: M3 output — reject-option classification when uncertainty > 0.15 + human-in-the-loop mandatory. Post-mitigation D5: 0.014.
SIGIL receipt: T93-fia-fm2-disability-counterfactual-rectified.
What happened: v0.3 of a content-moderation model under-flagged religious-minority hate speech (G5) due to lower training prevalence. The under-flagging caused downstream amplification (fewer removals → more exposure → more training data on the same under-represented pattern). D13 feedback-loop test flagged at month 6 with 0.18 spread.
Detection: T12 feedback-loop stability test on quarterly re-run.
Mitigation: M1 data — synthetic augmentation for G5 groups (validated per-group) + M4 governance — quarterly re-test + auto-rollback if spread > 0.10. Post-mitigation D13: 0.04 spread.
SIGIL receipt: T93-fia-fm3-feedback-loop-amplification-rectified.
Every DEFONEOS deployment ships with this annex auto-populated. The annex records:
vercel.app/defoneos-mod-fairness-impact-assessmentdefoneos-system-card.