# Sources, Integrations, and Evidence Foundation — Technical reference This reference owns the implementation contracts of `BND-EVIDENCE`. [Read the package guide](/documentation/packages/bnd-evidence) for usage, configuration, dependencies and upgrades. Generic execution, isolation and admission remain governed by [Platform Documentation](/documentation/platform). ## Data model and interfaces **DataModelPackage:** `DM-EVIDENCE` Owned entities: source connection, credential reference, source record, media, document version, evidence fragment, provenance, import job, object mapping. **InterfacePackage contributions:** `IF-EVIDENCE-ADMIN`, `IF-MEETING-CONTEXT`. 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-EVIDENCE-FOUNDATION` - Connector Authorization - Source Synchronization - Evidence Ingestion - Integrity and Provenance - Deduplication and Versioning - Source Correction - Dependency Invalidation - Business Object Resolution - Connector Recovery 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-SOURCE-ACCESS`, `EVAL-INGESTION-PROVENANCE`, calendar, email, document, file, CRM, project, webhook, and API connector packages. 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 ### P02 — Entity and Business Object Resolution Agent **Agent ID:** `effority.entity-object-resolution` **Purpose:** Resolves speakers, names, aliases, meetings, topics, documents, and findings to canonical people, accounts, opportunities, projects, deliverables, requirements, and other business objects. **Triggers:** Transcript publication, document ingestion, business-object update, speaker correction, and unresolved-object review. **Inputs:** Authorized names, metadata, participant lists, aliases, identifiers, topic context, graph relationships, and validity dates. **Outputs:** Candidate and confirmed entity links, multi-object assignments, confidence, alternatives, and unresolved cases. **Tools:** `directory.read`, `graph.read`, `business-object.search`, `calendar.metadata_read`, and `speaker-resolution.read`. **Review and limits:** Low-margin matches remain unresolved. The agent never creates a person or merges identities without the configured approval workflow. **Primary metrics:** Entity-linking accuracy, ambiguous-resolution rate, false merge rate, object-assignment accuracy, and review correction rate. ## Implementation profiles ### Document Ingestion and Knowledge Preparation The `document-service` converts email, calendars, PDFs, office files, CRM, and project managers into a canonical schema: - text, tables, and hierarchical structure; - source page, section, paragraph, cell, or clause; - author, date, version, validity, and source system; - entities and business objects; - ACL and sensitivity classification; - content hash and duplicate or citation relationships. Scanned PDFs pass through local OCR. Email threads are split by message and repeated quoted text is removed without losing the link to source. Contracts and SOWs are chunked by clause; plans by milestone, activity, or risk; transcripts by turn and topic. Every chunk retains enough parent context to prevent isolated interpretation. ### Multimodal Implementation When a source includes video or screen sharing, the `multimodal-service`: 1. detects scene changes and extracts timestamped keyframes; 2. removes redundant frames using perceptual hashing; 3. runs local OCR over slides, whiteboards, and screens; 4. classifies the visual-content type; 5. uses a local vision-language model to describe only relevant elements; 6. links spoken references—“this figure,” “the diagram,” “the date on the slide”—to the candidate frame; 7. indexes text, description, and coordinates under the media’s ACL; 8. compares displayed demos, timelines, or diagrams with the current proposal, SOW, and documents. The service does not perform facial recognition, emotion reading, gaze tracking, or physical-attribute inference. Frames containing credentials, personal data, or unauthorized windows are redacted before indexing. ### Local Automatic Translation Whisper always preserves the original-language transcript. For languages and destinations not covered by its translation task, the `translation-service` uses a local multilingual NMT model. - Translates by segment with topic context. - Applies a terminology glossary with constraints. - Maintains sentence-level alignment and timestamps. - Preserves names, figures, units, and codes through verified placeholders. - Performs a back-check of figures, negations, and critical terms. - Marks low confidence and supports human correction. - Never replaces original text as primary evidence. Summaries are generated directly in the target language from original or translated evidence; the report records the route used. ### AI-Assisted Redaction and Minimization Before indexing or sending context to a model, the privacy pipeline combines rules, NER, and tenant dictionaries to detect credentials, identifiers, financial information, health information, secrets, and configured categories. Policy decides whether to: - retain the data within the authorized environment; - replace it with a reversible token kept in a separate vault; - redact it irreversibly; - block the excerpt; - restrict it to a specific model or role. The analysis model receives tokens such as `[PERSON_3]` or `[ACCOUNT_NUMBER_1]` when the real value is unnecessary. Rehydration occurs only after output validation and a new permission check. Tokens are not global or reusable across tenants. --- ## Integration boundaries An imported document is evidence, not a verified business conclusion. Email and document findings are reviewed through BND-FINDINGS, just like meeting findings. Owned screen contracts: [SCR-MTG-05](/documentation/packages/bnd-evidence/screens/home-meetings-context), [SCR-MTG-06](/documentation/packages/bnd-evidence/screens/home-meetings-files-and-versions), [SCR-MTG-07](/documentation/packages/bnd-evidence/screens/home-meetings-activity), [SCR-ADM-03](/documentation/packages/bnd-evidence/screens/administration-sources-and-integrations). 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.