Information Sharing Architecture¶

DueCare is built around a simple rule: sensitive material is processed where it is owned, while only reviewed and anonymized intelligence can move between deployments.

Purpose¶

The architecture lets many local nodes improve shared knowledge without centralizing raw worker case data. A node can be a Kaggle session, an NGO office machine, a platform safety deployment, a regulator workstation, a mobile client, or a researcher environment. Each node can run Gemma 4, harness layers, document review, graph extraction, search safety, and anonymization locally.

When a user chooses to share, DueCare converts reviewed local outputs into portable objects: sanitized facts, evidence edges, public-source proposals, aggregate signals, hash receipts, benchmark rows, or pack updates.

Local Node Responsibilities¶

Responsibility	What Happens Locally
Review source material	The node handles messages, documents, screenshots, contracts, receipts, IDs, and case notes.
Generate drafts	Gemma 4 and the safety guidance layer produce summaries, questions, citations, risk notes, and graph edges.
Verify evidence	Reviewers inspect citations, evidence quotes, uncertainty, and graph links before promotion.
Anonymize	The local workflow redacts PII and converts approved outputs into sanitized objects.
Decide consent	The submitting user or organization chooses what can be shared, attributed, aggregated, or kept private.

Shared Object Types¶

Object	Use
`CasePatternObservation`	Sanitized pattern such as fee pressure, document control, threats, false promises, or repeated recruiter language.
`EvidenceEdge`	Reviewed relationship between entities, documents, payments, events, locations, or journey stages.
`RegulationUpdate`	Public-source legal or policy update with source URL, jurisdiction, effective date, and review status.
`CorridorFeeRecord`	Public or vetted fee rule for a corridor, sector, visa path, or recruitment stage.
`ContactMetadataUpdate`	Consented or public-source update to contact metadata, never raw private contact logs.
`BenchmarkRow`	Anonymized or synthetic prompt/evidence row for reproducible model and harness evaluation.
`PackProposal`	Proposed change to a versioned knowledge pack, with provenance and curator notes.

Hub Responsibilities¶

The public hub coordinates shared intelligence. It should store public-source facts, vetted pack metadata, anonymized aggregate signals, hash receipts, curation status, and consented contact metadata.

It should not store raw worker chats, phone numbers, addresses, passports, private documents, unredacted screenshots, or personal case narratives.

Hub Function	Purpose
Intake validation	Reject payloads that look like raw PII or private case narratives.
Review queue	Hold sanitized proposals until a curator or trusted reviewer approves them.
Pack registry	Publish versioned knowledge packs that local nodes can pull and verify.
Audit receipts	Record hashes, timestamps, source references, and review state.
Aggregate signals	Show corridor, sector, and pattern-level trends only after threshold and consent checks.
Evaluation artifacts	Publish benchmark rows, scorecards, and reproducibility metadata.

Step	Description
1. Local review	A local node reviews source material and builds drafts, graph edges, citations, and risk notes.
2. Human approval	A reviewer decides which outputs are accurate enough to promote.
3. Local anonymization	PII and private narrative details are removed before anything leaves the node.
4. Submission	The node submits a sanitized object, aggregate signal, public-source proposal, or hash receipt.
5. Server validation	The hub rejects raw PII and routes accepted objects into a review queue.
6. Curation	Curators approve, reject, merge, or request more evidence.
7. Pack update	Approved objects become knowledge-pack updates, benchmark rows, or aggregate trend data.
8. Reuse	Other nodes pull the updated packs and benefit from the improved shared intelligence.

User Mapping¶

User Path	Information Sharing Role
Platform safety	Shares anonymized abuse-pattern signals and public policy updates from review queues.
NGO & regulator	Shares reviewed, sanitized case patterns and corridor evidence without uploading raw case files.
Individual worker / mobile	Shares only worker-approved notes, intake drafts, or anonymized signals.
Researcher	Shares reproducible benchmark artifacts, source audits, and evaluation metadata.
Anonymized knowledge sharing	Operates the redaction, consent, curation, and pack-proposal path.
Developer / integration partner	Embeds the same object schemas and validation boundaries into external systems.