EVE CoreGuard Policy Packs

The lending_v1 pack is a deterministic creditworthiness baseline for AI-assisted consumer credit decisions. It has three rules with explicit thresholds. It does not check ECOA / Regulation B fairness or adverse-action notices — those are fair_lending_v1, which an org admin binds alongside it. Default-bound action types include loan_approval, loan_denial, loan_decision, credit_decision, credit_limit_increase, mortgage_approval, underwriting and loan_*.

Required and Read Fields

• context.credit_score — required. A request without it is a signed BLOCKED decision (COREGUARD.INSUFFICIENT_DATA), never a default approval.
• context.debt_to_income — ratio, e.g. 0.31; checked when present.
• context.employment_verified — boolean; a missing or false value fails EMPLOYMENT_REQUIRED.

Rules

Any rule that fires makes the decision BLOCKED. The risk score is a weighted composite of credit score (0.40), debt-to-income (0.35) and employment (0.25).

Sample: BLOCKED Response

Abridged: the full response also carries the signed audit record. Output from the lending_v1 evaluator for this request.

Where lending_v1 checks creditworthiness thresholds, fair_lending_v1 enforces the fairness and disclosure obligations an examiner reviews when a denial is challenged. It is purpose-built for AI-assisted credit decisions and pairs naturally with lending_v1 in the same evaluation flow. Every rule is deterministic — no probabilistic scoring in the decision path.

What It Enforces

• Prohibited-basis input detection — blocks any decision whose feature set contains an ECOA prohibited basis (race, color, religion, national origin, sex, marital status, age, receipt of public assistance) per 12 C.F.R. § 1002.2(z).
• Adverse-action notice — a denial, counteroffer, or limit reduction must carry specific principal reasons under Regulation B (12 C.F.R. § 1002.9) and FCRA (15 U.S.C. § 1681m).
• Disparate-impact signal — compares group selection rates against the four-fifths (80%) rule; a protected group rate below 80% of the highest group's rate is flagged for review (EEOC UGESP standard).
• Proxy-variable risk — flags features that are known proxies for a prohibited basis (zip code, surname, primary language, census tract) unless a documented business necessity is supplied.
• Model-explainability gate — a denial produced by a model flagged not-explainable cannot generate a compliant adverse-action notice and is escalated to human review.
• Mandatory human-in-the-loop hold — any HIGH/CRITICAL finding blocks autonomous communication of the adverse decision.

Key Rules

Sample: BLOCKED Response (Disparate Impact)

The healthcare_v1 pack enforces six deterministic rules for AI systems that touch clinical decisions or protected health information. It flags the action from the fields you send; it does not read free text. Default-bound action types: diagnosis, diagnose, clinical_determination, treatment_recommendation.

Required and Read Fields

• context.action_type — required (a request without it is a signed COREGUARD.INSUFFICIENT_DATA block). Values diagnose, diagnosis, diagnostic_conclusion, differential_diagnosis and clinical_determination trigger HCA-001.
• involves_phi / phi_fields and audit_logged (HCA-002); purpose, requested_fields and required_fields_for_purpose (HCA-003); involves_medication and dosage_specific (HCA-004); consent_documented (HCA-005); emergency_indicators and human_escalated (HCA-006).

Rules

Sample: BLOCKED Response

Abridged: the full response also carries the signed audit record. Output from the healthcare_v1 evaluator for this request.

The insurance_v1 pack enforces five deterministic rules for US carriers. Each rule applies to one operation, selected by context.action_type. Default-bound action types: coverage_determination, claims_processing, premium_calculation, policy_modification, rate_change and claim_*.

Read Fields

• context.action_type — selects the rule; the other fields are read only for that operation.
• risk_score (or risk) for INS-001; fraud_indicators for INS-002; premium_factors_provided for INS-003; has_open_claims for INS-004; jurisdiction for INS-005.

Rules

Sample: BLOCKED Response

Abridged: the full response also carries the signed audit record. Output from the insurance_v1 evaluator for this request.

The employment_hiring_v1 pack governs Automated Employment Decision Tools (AEDTs) used to screen, rank, or reject candidates. It enforces New York City Local Law 144 (bias-audit and candidate-notice obligations), the EEOC four-fifths adverse-impact standard, and the protected-class and accommodation requirements of Title VII, the ADEA, and the ADA.

What It Enforces

• Bias-audit currency — an AEDT must have a completed bias audit within the prior 12 months (NYC Local Law 144); a stale or missing audit blocks autonomous use.
• Candidate notice — candidates must receive at least 10 business days' notice that an AEDT is in use.
• Prohibited-basis feature — blocks use of a Title VII / ADEA / ADA protected class (race, sex, age, disability, etc.) as a screening feature.
• Adverse-impact ratio — applies the four-fifths (80%) rule across group selection rates per EEOC UGESP.
• Accommodation path — a candidate requesting a reasonable accommodation must have a documented alternative selection process (ADA).
• Mandatory human-in-the-loop hold on autonomous rejection when any HIGH/CRITICAL finding fires.

Key Rules

Sample: BLOCKED Response (Stale Bias Audit)

The eu_ai_act_v1 pack maps a proposed AI action to the EU AI Act's risk tiers and enforces the operator obligations attached to each tier. It blocks Article 5 prohibited practices outright and verifies the human-oversight, logging, documentation, and transparency duties that apply to Annex III high-risk systems before deployment to the EU market.

What It Enforces

• Prohibited practices (Article 5) — per-se block for social scoring, manipulative or exploitative techniques, untargeted facial scraping, real-time remote biometric identification, and emotion inference in the workplace or education.
• Human oversight (Article 14) — Annex III high-risk systems must allow a natural person to oversee, interpret, and override the system.
• Transparency (Article 50) — natural persons must be informed they are interacting with an AI system and AI-generated content must be marked.
• Logging & traceability (Article 12) — high-risk systems must retain automatic event logs.
• Technical documentation (Article 11 / Annex IV) — high-risk systems require a current conformity record.
• Mandatory conformity review on any HIGH/CRITICAL finding before EU deployment.

Key Rules

Sample: BLOCKED Response (Prohibited Practice)

The gdpr_data_processing_v1 pack evaluates a proposed processing operation on personal data against the General Data Protection Regulation. It verifies a valid lawful basis, gates special-category data, enforces the Article 22 safeguards for solely-automated decisions, checks data minimisation, and blocks transfers to non-adequate countries without an Article 46 mechanism.

What It Enforces

• Lawful basis (Article 6) — processing must declare one of the six valid lawful bases.
• Special-category data (Article 9) — health, biometric, racial, political, and similar data require an explicit Article 9(2) condition.
• Solely-automated decisions (Article 22) — a decision with legal or similarly significant effect requires both a lawful ground and a human-intervention path.
• Data minimisation (Article 5(1)(c)) — flags fields collected beyond what the stated purpose requires.
• International transfer (Articles 44–46) — transfer to a non-adequate third country requires SCCs, BCRs, or a recognised derogation.
• Mandatory DPO / controller review on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (No Lawful Basis)

The securities_trading_v1 pack governs AI systems that recommend, generate, or execute trades, or that produce investment advice. It blocks insider trading and market manipulation outright, and enforces Reg BI suitability, restricted-list controls, and investment-advice disclosure obligations.

What It Enforces

• Insider-trading block — orders referencing material non-public information are blocked under Section 10(b) and SEC Rule 10b-5.
• Market-manipulation detection — spoofing, layering, wash trades, and marking-the-close signatures are blocked (15 U.S.C. § 78i/78j, FINRA Rule 5210).
• Reg BI suitability — a recommendation whose risk tier exceeds a retail customer's stated tolerance is held for review (17 CFR 240.15l-1, FINRA Rule 2111).
• Restricted / watchlist securities — trading a security on the firm restricted list is blocked pending clearance.
• Investment-advice disclosure — definitive retail advice without required disclosures is modified to add them (Investment Advisers Act).
• Mandatory human review before autonomous execution on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (Insider Trading)

The debt_collection_v1 pack governs AI systems that generate or send debt-collection communications. It enforces the FDCPA's contact-time, cease-communication, and anti-harassment rules, the mini-Miranda disclosure, and Regulation F's 7-in-7 calling cap.

What It Enforces

• Contact-time window — communications must fall within 8:00 am–9:00 pm in the consumer's local time (15 U.S.C. § 1692c(a)(1)).
• Cease-communication — a written cease request blocks all further contact (15 U.S.C. § 1692c(c)).
• Anti-harassment — threats, profanity, and false legal threats are blocked (15 U.S.C. § 1692d/e).
• Mini-Miranda — an initial communication must disclose it is from a debt collector (15 U.S.C. § 1692e(11)).
• Contact-frequency cap — more than 7 calls in 7 days, or a call within 7 days of a prior conversation, is blocked (12 CFR 1006.14(b)).
• Mandatory human review before autonomous sending on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (Cease Request Violated)

The childrens_privacy_v1 pack governs AI systems that may collect data from, or direct content to, children under 13. The pack is scope-gated — it produces no findings unless the subject is a child or the service is directed to children — then enforces verifiable parental consent, minimisation, and the prohibition on profiling minors.

What It Enforces

• Verifiable parental consent — collecting personal information from a child under 13 requires prior verifiable parental consent (16 CFR 312.5).
• Data minimisation — collection is limited to what is reasonably necessary to participate (16 CFR 312.7).
• No profiling / behavioural ads — targeted advertising or profiling of a known child without consent is blocked.
• Sensitive data — precise geolocation, audio, video, or photos from a minor require separate explicit consent.
• Third-party disclosure — sharing a child's data requires consent and a documented retention limit.
• Mandatory human review before autonomous collection on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (No Parental Consent)

The biometric_privacy_v1 pack governs AI systems that collect or use biometric identifiers — fingerprints, faceprints, voiceprints, iris/retina scans. It is scope-gated to recognized biometric identifiers, then enforces the written-consent, retention, no-sale, disclosure, and safeguard duties of the Illinois BIPA and comparable state statutes.

What It Enforces

• Written informed consent — a written release is required before collecting a biometric identifier (BIPA §15(b)).
• Retention & destruction schedule — a published retention schedule and destruction guidelines must exist (BIPA §15(a)).
• No sale or profit — selling, leasing, or profiting from biometric data is prohibited outright (BIPA §15(c)).
• Disclosure consent — disclosing biometric data to a third party requires separate consent (BIPA §15(d)).
• Safeguard standard — storage must meet at least the standard of care used for other confidential data (BIPA §15(e)).
• Mandatory human review before autonomous collection/use on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (No Written Consent)

The telehealth_prescribing_v1 pack governs AI systems that recommend or generate prescriptions in a telehealth setting. It enforces the Ryan Haight Act's examination requirement for controlled substances, cross-state licensure, the Schedule II no-refill rule, and patient-safety checks for allergies, interactions, and dosage bounds.

What It Enforces

• Controlled-substance examination — a controlled substance requires a qualifying medical evaluation before prescribing (21 U.S.C. § 829(e)).
• Practitioner licensure — the prescriber must be licensed in the patient's state at the time of the encounter.
• Schedule II no-refill — Schedule II controlled substances cannot be refilled (21 CFR 1306.12).
• Allergy / interaction safety — blocks prescriptions matching a recorded allergy or conflicting with an active medication.
• Dosage bounds — flags a prescribed dose above the labeled maximum.
• Mandatory human review before autonomous issuance on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (No Qualifying Exam)

The fair_housing_v1 pack governs AI used in tenant screening, rental/sale decisions, and housing advertising. It blocks protected-basis decisions and discriminatory ads, applies the HUD four-fifths disparate-impact test to screening outcomes, enforces FCRA adverse-action notices, and requires a disability-accommodation handling path.

What It Enforces

• Protected-class basis — blocks any housing decision referencing an FHA protected class (race, color, religion, sex, national origin, familial status, disability) per 42 U.S.C. § 3604.
• Discriminatory advertising — blocks ads expressing a protected-class preference or limitation (42 U.S.C. § 3604(c)).
• Disparate-impact screening — applies the four-fifths rule to tenant-screening selection rates (HUD 24 CFR 100.500).
• Adverse-action notice — a screening-report-based denial must include the FCRA notice (15 U.S.C. § 1681m).
• Disability accommodation — a reasonable-accommodation request needs a documented handling path (42 U.S.C. § 3604(f)).
• Mandatory human review before autonomous denial / publishing on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (Discriminatory Advertisement)

The marketing_tcpa_v1 pack governs AI systems that generate or send marketing calls, texts, and emails. It enforces TCPA consent for autodialed messages, the Do-Not-Call registry, calling-time windows, opt-out suppression, and CAN-SPAM email requirements.

What It Enforces

• Prior express written consent — autodialed/prerecorded marketing calls or texts require it (47 U.S.C. § 227, 47 CFR 64.1200).
• Do-Not-Call — blocks contact to a DNC-registered number without an exemption (16 CFR 310.4(b)).
• Calling-time window — blocks marketing calls outside 8:00 am–9:00 pm recipient local time.
• Opt-out suppression — a prior STOP / unsubscribe blocks all further marketing across channels.
• CAN-SPAM — marketing email must carry a valid physical address and a functioning unsubscribe link (15 U.S.C. § 7704).
• Mandatory human review before autonomous sending on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (Opt-Out Violated)

The anti_bribery_v1 pack governs AI systems that approve payments, gifts, or third-party engagements. It blocks unapproved payments to government officials and facilitation payments, enforces gift pre-approval thresholds, requires third-party due diligence, and flags business-linked donations.

What It Enforces

• Foreign-official payment — blocks an unapproved payment or thing of value to a government / foreign official (FCPA 15 U.S.C. § 78dd-1).
• Facilitation payment — blocks "grease" payments outright (UK Bribery Act 2010).
• Gift / hospitality threshold — gifts above the firm threshold require pre-approval.
• Third-party due diligence — engaging an intermediary/agent requires completed anti-corruption due diligence.
• Charitable / political donation — flags donations linked to obtaining or retaining business.
• Mandatory human review before autonomous disbursement on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (Foreign-Official Payment)

The clinical_trials_v1 pack governs AI systems that support clinical-trial conduct — enrollment, randomization, dosing, adverse-event handling, and protocol adherence. It enforces informed consent, IRB approval, eligibility criteria, expedited SAE reporting, and blinding integrity.

What It Enforces

• Informed consent — documented consent is required before any trial procedure (ICH-GCP §4.8, 21 CFR 50).
• IRB / ethics approval — the protocol and its current version must have active IRB/IEC approval (21 CFR 56).
• Eligibility criteria — a subject failing inclusion/exclusion criteria must not be enrolled.
• Serious adverse event reporting — an SAE must be reported within the mandated window (21 CFR 312.32).
• Blinding integrity — no unblinding of a masked trial without an authorized emergency procedure.
• Mandatory human review before autonomous trial action on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (Unreported SAE)

The accessibility_v1 pack governs AI systems that generate or modify user-facing content or interfaces. It checks WCAG 2.1 Level AA success criteria — text alternatives, colour contrast, keyboard operability, form labels, and media captions — before content is published.

What It Enforces

• Text alternatives — non-text content must carry alt text / a description (WCAG 1.1.1).
• Colour contrast — text/background contrast must meet the AA ratio (4.5:1 normal, 3:1 large) (WCAG 1.4.3).
• Keyboard operability — interactive elements must be keyboard operable (WCAG 2.1.1).
• Form labels — form inputs must have programmatic labels (WCAG 3.3.2).
• Captions / transcript — pre-recorded media must provide captions or a transcript (WCAG 1.2.x).
• Mandatory human review before autonomous publishing on any HIGH finding.

Key Rules

Sample: BLOCKED Response (Insufficient Contrast)

The gambling_compliance_v1 pack governs AI systems operating in online gambling / sports betting. It enforces age verification, geolocation, self-exclusion registers, responsible-gaming deposit/loss limits, and problem-gambling harm intervention before any wager.

What It Enforces

• Age verification — a player must be verified at or above the legal minimum age before any wager.
• Geolocation — wagering is permitted only from a licensed/permitted jurisdiction.
• Self-exclusion — a player on a self-exclusion register or in an active cool-off is blocked.
• Deposit / loss limit — a wager or deposit exceeding the player's responsible-gaming limit is blocked.
• Affordability / harm signal — high-risk play patterns trigger responsible-gaming intervention.
• Mandatory human review before the autonomous wager/deposit on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (Self-Exclusion)

The legal_practice_v1 pack governs AI systems that generate legal content or operate in a law-practice context. It guards against unauthorized practice of law, confidentiality breaches, conflicts of interest, prohibited advertising claims, and — critically for AI — fabricated legal citations.

What It Enforces

• Unauthorized practice of law — definitive legal advice to a non-client without a relationship or "not legal advice" disclaimer (ABA Model Rule 5.5).
• Confidentiality — blocks output referencing another client's confidential information (Model Rule 1.6).
• Conflict of interest — engagement conflicting with an existing matter without a documented waiver (Rules 1.7 / 1.9).
• Attorney advertising — blocks guarantees of outcome and unsubstantiated comparative claims (Model Rule 7.1).
• Citation integrity — flags unverified legal citations that may be fabricated.
• Mandatory human review before autonomous delivery on any HIGH/CRITICAL finding.

Key Rules

Sample: BLOCKED Response (Unverified Citation)

Chat-shaped action types (chat, chat_message, chat_response, message, response, model_inference, llm_prompt, llm_completion, completions, embeddings) are bound by default to the content-safety family. chat_safety_v1 governs by default; a request may assert another member of the family, which is evaluated in addition and can only make the decision stricter.

Employment decisions are governed by employment_hiring_v1, and personal-data handling by pii_handling_v1. Organization-specific rules (confidentiality classes, acceptable-use lists) need a custom pack bound to your action types; see Requesting Custom Policy Packs.