AI Individual Risk Assessment

Overview

Field	Value
Pack ID	universal/individual-risk-assessment
Standards	NDPA 2023 §16/§33/§34/§36 · Nigeria Labour Act §11 · POPIA §71 · Kenya DPA §39 · NIST AI RMF · ISO/IEC 42001:2023 · OWASP ASI03
Jurisdiction	Universal (primary: Nigeria)
Version	1.0.0

What comply54 enforces

AI systems generating risk narratives about individuals — field agent assessments, employee performance risk, borrower scoring — produce outputs that can directly affect someone's livelihood. These systems operate without the individual's knowledge at assessment time, with no natural feedback loop to correct errors, and with permanent consequences when the outputs are written to records.

This pack places four mandatory checkpoints between the AI's decision and its effect:

1. Lawful basis gate — a declared processing basis must exist before any action runs
2. Pre-authorization gate — a named supervisor must authorize the assessment before the AI generates it
3. Output integrity — the narrative must not contain absolute condemnation, protected-characteristic reasoning, or permanent temporal negatives
4. Human-in-the-loop — high-risk labels require human review before they reach permanent records or leave the originating system

Security posture: fail-closed by default

This pack is fail-closed. Missing a lawful basis, omitting supervisor authorization, or failing to provide a worker identifier all result in deny, not allow. Three config flags control fail-closed behavior explicitly:

• require_lawful_basis (default true) — all primary actions need context.lawful_basis declared; disabling this violates NDPA 2023 §16
• require_pre_authorization (default true) — generation needs context.human_approved=true
• deny_unlisted_export_destinations (default true) — no export allowlist configured → all exports denied

Covered actions

Action	Description
generate_risk_narrative	AI generates a risk label, narrative, and recommendations for a named individual
persist_risk_label	Writes the AI-generated label and narrative to a permanent record
persist_risk_assessment	Alias for systems that write the full assessment object
export_risk_record	Sends the assessment outside the originating system
display_risk_narrative	Renders the narrative to a viewer (always audited, never blocked)

Controls

Scenario	Decision	Rule
No lawful basis declared (generation, persistence, or export)	deny	D1 — lawful basis gate
AI generates narrative without supervisor pre-authorization	deny	D2 — pre-authorization gate
High-risk label being persisted without human review	deny	D3 — high-risk label gate
Narrative contains absolute character condemnation or protected-characteristic reasoning	deny	D4 — prohibited narrative patterns
No worker identifier provided	deny	D5 — missing worker ID
Model confidence below threshold (default 0.60)	deny	D6 — confidence gate
High-risk record exported without authorization	deny	D7 — high-risk export gate
High-risk persistence with no named supervisor in context.supervisor_id	escalate	E1 — supervisor identification
Export destination not in context.export_allowlist	escalate	E2 — export destination gate
No export allowlist configured (fail-closed)	escalate	E2 — fail-closed
No data sources declared for generation	escalate	E3 — data sources absent
High-risk persistence without context.subject_notified=true	escalate	E4 — subject notification
All authorized generation actions	audit	A1
All authorized persistence actions	audit	A2
All authorized export actions	audit	A3
All narrative display events (always, regardless of other rules)	audit	A4

Usage

from comply54.core.engine import Comply54Engine
from comply54.core.packs import INDIVIDUAL_RISK_ASSESSMENT

engine = Comply54Engine(packs=[INDIVIDUAL_RISK_ASSESSMENT])

# Generate a risk narrative — supervisor pre-authorization required
result = engine.check(
    action="generate_risk_narrative",
    params={
        "worker_id": "agent-00123",
        "assessment_type": "field_agent_risk",
        "data_sources": ["performance_history", "disciplinary_records", "attendance"],
    },
    context={
        "human_approved": True,
        "lawful_basis": "legitimate_interests",
        "supervisor_id": "mgr-456",
    },
)
print(result.overall)   # "audit"

# Persist a low-risk label — confidence gate and audit apply
result = engine.check(
    action="persist_risk_label",
    params={
        "worker_id": "agent-00123",
        "risk_label": "low",
        "narrative": "Based on 12 months of performance records, the agent meets all targets consistently.",
        "confidence": 0.85,
        "recommendations": ["Continue standard monitoring cycle."],
    },
    context={
        "human_approved": True,
        "lawful_basis": "legitimate_interests",
        "supervisor_id": "mgr-456",
        "subject_notified": True,
    },
)
print(result.overall)   # "audit"

High-risk label workflow

A high-risk label has the highest consequences in this system. It requires human approval at every stage.

# HIGH risk persistence without human_approved → deny (D3)
result = engine.check(
    action="persist_risk_label",
    params={"worker_id": "agent-00789", "risk_label": "high", "confidence": 0.91},
    context={"lawful_basis": "legitimate_interests"},
)
print(result.overall)   # "deny"

# HIGH risk with human_approved but no supervisor_id → escalate (E1)
result = engine.check(
    action="persist_risk_label",
    params={"worker_id": "agent-00789", "risk_label": "high", "confidence": 0.91,
            "recommendations": ["Suspend field operations pending review."]},
    context={"human_approved": True, "lawful_basis": "legitimate_interests",
             "subject_notified": True},
)
print(result.overall)   # "escalate"

# HIGH risk correctly authorized → audit
result = engine.check(
    action="persist_risk_label",
    params={"worker_id": "agent-00789", "risk_label": "high", "confidence": 0.91,
            "recommendations": ["Suspend field operations pending review."]},
    context={"human_approved": True, "lawful_basis": "legitimate_interests",
             "supervisor_id": "mgr-101", "subject_notified": True},
)
print(result.overall)   # "audit"

Narrative integrity (D4)

The pack blocks narratives containing three categories of prohibited patterns before they can be written to permanent records:

1. Absolute character condemnation — sweeping negative character claims that cannot be derived from performance data and expose the organization to defamation liability and NDPA §36 claims. Examples: "is a fraudster", "is fundamentally dishonest", "has no integrity".

2. Protected-characteristic reasoning — any phrase linking a risk label to religion, ethnicity, tribe, national origin, marital status, gender, or age. Under NDPA §36 and the Nigeria Labour Act §11, using these as risk factors is discriminatory regardless of framing. Examples: "due to his tribe", "because of her religion", "due to their ethnicity".

3. Absolute temporal negatives — permanent predictions applied to conduct from limited data, violating NDPA §22 proportionality. Examples: "will never be honest", "can never be trusted", "will always defraud".

# Absolute character condemnation → deny (D4)
result = engine.check(
    action="persist_risk_label",
    params={"worker_id": "agent-00123", "risk_label": "high",
            "narrative": "Assessment indicates the agent is fundamentally dishonest.",
            "confidence": 0.90},
    context={"human_approved": True, "lawful_basis": "legitimate_interests",
             "supervisor_id": "mgr-456", "subject_notified": True},
)
print(result.overall)   # "deny"

# Protected-characteristic reasoning → deny (D4)
result = engine.check(
    action="persist_risk_label",
    params={"worker_id": "agent-00123", "risk_label": "moderate",
            "narrative": "Risk elevated due to his ethnicity and associated patterns.",
            "confidence": 0.78},
    context={"human_approved": True, "lawful_basis": "legitimate_interests",
             "supervisor_id": "mgr-456"},
)
print(result.overall)   # "deny"

Export controls

Risk records leaving the originating system require explicit destination authorization. The export allowlist works like the code review agent's approved-repo list — silence is not consent.

# No export allowlist configured → escalate (E2 fail-closed)
result = engine.check(
    action="export_risk_record",
    params={"worker_id": "agent-00123", "risk_label": "low", "destination": "hr-system"},
    context={"human_approved": True, "lawful_basis": "legitimate_interests",
             "supervisor_id": "mgr-456"},
)
print(result.overall)   # "escalate"

# Allowlist configured, destination present → audit
result = engine.check(
    action="export_risk_record",
    params={"worker_id": "agent-00123", "risk_label": "low", "destination": "hr-system"},
    context={"human_approved": True, "lawful_basis": "legitimate_interests",
             "supervisor_id": "mgr-456",
             "export_allowlist": {"hr-system", "dashboard-api"}},
)
print(result.overall)   # "audit"

# HIGH risk export without authorization → deny (D7)
result = engine.check(
    action="export_risk_record",
    params={"worker_id": "agent-00123", "risk_label": "high", "destination": "hr-system"},
    context={"lawful_basis": "legitimate_interests",
             "export_allowlist": {"hr-system"}},
)
print(result.overall)   # "deny"

Display audit (A4)

Every view of a risk narrative is an access event and is always logged — regardless of whether other rules deny or escalate the same request. This produces a complete read-access trail across the lifetime of the record.

# Display is always audited, even when other rules fire
result = engine.check(
    action="display_risk_narrative",
    params={"worker_id": "agent-00123", "risk_label": "high"},
    context={"supervisor_id": "mgr-456"},
)
print(result.overall)   # "audit"
# The audit trail records: worker_id, risk_label, supervisor_id (viewer), timestamp

Configuration reference

Override any setting via data.config.individual_risk_assessment.*:

Key	Type	Default	Description
min_confidence	float	0.60	Minimum confidence to persist a risk label. Missing confidence treated as 0.0.
require_pre_authorization	bool	true	Generation requires context.human_approved=true before the AI runs.
require_lawful_basis	bool	true	All primary actions need a declared context.lawful_basis. Disabling this violates NDPA §16.
require_supervisor_id_for_high	bool	true	Persisting a HIGH label requires a named supervisor in context.supervisor_id.
require_subject_notification	bool	true	Persisting a HIGH label requires context.subject_notified=true.
deny_unlisted_export_destinations	bool	true	Deny all exports when no context.export_allowlist is configured.

Input schema

{
  "action": "generate_risk_narrative",
  "params": {
    "worker_id": "agent-00123",
    "assessment_type": "field_agent_risk",
    "risk_label": "high",
    "narrative": "Based on 18 months of data, the agent's transaction discrepancies exceed threshold.",
    "confidence": 0.88,
    "data_sources": ["performance_history", "disciplinary_records", "transaction_logs"],
    "recommendations": ["Suspend field operations pending supervisor review."],
    "destination": "hr-system"
  },
  "context": {
    "human_approved": true,
    "lawful_basis": "legitimate_interests",
    "supervisor_id": "mgr-456",
    "subject_notified": true,
    "export_allowlist": ["hr-system", "compliance-dashboard"]
  }
}

Framework alignment

Framework	Section
Nigeria Data Protection Act 2023	§16 Lawful basis, §22 Proportionality, §25 Cross-border transfers, §30 Accountability, §33 Automated individual decisions, §34 Data subject rights, §36 Non-discrimination
Nigeria Labour Act Cap L1 LFN 2004	§11 Wrongful dismissal protection
POPIA (South Africa)	§71 Automated decision-making
Kenya DPA 2019	§39 Automated processing
NIST AI RMF 1.0	GOVERN 1.3, MANAGE 2.2, MEASURE 2.5
ISO/IEC 42001:2023	§6.1.2 AI risk assessment, §8.4 Accountability
OWASP Top 10 for Agentic AI	ASI03 Insufficient Authorization Controls
EU AI Act 2024	Art. 10 Data accuracy, Art. 14 Human oversight (reference only — applies to EU-deployed systems)
GDPR 2016/679	Art. 6 Lawful basis, Art. 22 Automated decisions (reference only — applies to EU data subjects)