# Ethics and Compliance Review — Technical reference

This reference owns the implementation contracts of `BND-ETHICS`. [Read the package guide](/documentation/packages/bnd-ethics) for usage, configuration, dependencies and upgrades. Generic execution, isolation and admission remain governed by [Platform Documentation](/documentation/platform).

## Data model and interfaces

**DataModelPackage:** `DM-ETHICS`

Owned entities: candidate alert, restricted case, policy mapping, reviewer assignment, case decision, safeguarded aggregate, disclosure record.

**InterfacePackage contributions:** `IF-ETHICS-RESTRICTED`.

Each business object and source reference remains tenant- and purpose-scoped. The package owns schema evolution and declared migration behavior; the platform enforces schema registration, authorization and versioned installation. Consumers reference this model rather than creating competing canonical copies.

## Workflows

**WorkflowPackage:** `WFP-ETHICS-REVIEW`

- Candidate Alert Detection
- Segregated Review Routing
- Case Opening
- Policy Mapping
- Restricted Evidence Review
- Case Decision
- Authorized Escalation
- Safeguarded Aggregation
- Reviewer Audit Reconstruction

These are named definitions within the workflow package, not separate invented workflow IDs. Workflow runs resolve package versions, actor and purpose, authorized inputs, retries, review gates and output schemas through the environment lock. Temporal provides durable execution and Kafka carries declared lifecycle/domain events.

## Policies, evaluations and tools

`POL-ETHICS-SEGREGATION`, `POL-ETHICS-DISCLOSURE`, `EVAL-ETHICS-HIGH-PRECISION`, restricted case and disclosure tools.

Qualification covers the selected model/runtime profile, languages, access scope, failure handling and the package's domain-specific outcomes. Tool execution uses the platform gateway; package code cannot increase its own permissions or write directly to external systems.

## Owned agents

### B05 — Ethics and Compliance Agent

**Agent ID:** `effority.ethics-compliance`

**Purpose:** Produces high-precision candidate alerts for possible inappropriate language, harassment, threat, discrimination, conflict, bribery, retaliation, disclosure, or other configured policy concerns.

**Triggers:** Eligible transcript or communication, policy update, reviewer correction, or restricted reprocessing request.

**Inputs:** Minimum necessary context window, configured policy, cultural and language profile, difficult counterexamples, speaker and target context, and source quality.

**Outputs:** Candidate category, neutral statement, target, subject, severity, confidence, evidence, alternatives, and segregated review requirement.

**Tools:** `restricted_evidence.read`, `policy.read`, `citation.verify`, `case.read_authorized`, and `review.route_restricted`.

**Review and limits:** A candidate alert is not an allegation or established fact. E2–E4 findings enter segregated human review. The agent never sanctions, diagnoses, or infers protected attributes.

**Primary metrics:** Category precision and recall, severe false-negative rate, false-positive rate, reviewer rejection rate, language-segment quality, and time to safety routing.



## Implementation profiles

### Ethics-Alert Implementation

The ethics module uses a high-precision cascade separated from general analysis:

1. rules and lightweight classifiers generate high-recall candidates;
2. a contextual classifier removes quotations, negations, readings, and nondirected uses;
3. a local LLM receives only the necessary window, applicable policy, and counterexamples;
4. a verifier checks citations, subject, target, category, and severity;
5. E2–E4 alerts enter a segregated review queue;
6. only a human decision opens, dismisses, or escalates a formal case.

Severity and confidence are independent fields: a signal can be potentially serious while also having low confidence. Low-confidence alerts are never promoted automatically. The system does not infer emotions, personality, health, or protected attributes.

Testing for this module includes difficult negatives, cultural variation, irony, third-party discussion, educational examples, jokes, complaints, technical language, and incorrectly transcribed audio. Thresholds are configured by language; a category without sufficient evaluation remains technically disabled for that language.




## Integration boundaries

Alerts are candidates for review, not findings of guilt. Private personal-development recommendations are governed separately by BND-DEVELOPMENT.

Owned screen contracts: [SCR-ETH-01](/documentation/packages/bnd-ethics/screens/intelligence-ethics-review-queue), [SCR-ETH-02](/documentation/packages/bnd-ethics/screens/intelligence-ethics-cases), [SCR-ETH-03](/documentation/packages/bnd-ethics/screens/intelligence-ethics-policies), [SCR-ETH-04](/documentation/packages/bnd-ethics/screens/intelligence-ethics-restricted-analytics), [SCR-ETH-05](/documentation/packages/bnd-ethics/screens/intelligence-ethics-reviewer-audit).

See [solution workflows](/documentation/technology-consulting-delivery/connected-workflows) for cross-package behavior. Screens, agents and workflows keep one canonical owner even when other bundles invoke or enrich them.