Nava Labs Field Notes

How our SNAP evaluation led to a fix in Inspect AI

Why GPT-5.6 models failed on Amazon Bedrock, and how a one-change fix became four in the UK AI Security Institute’s eval framework.

Oct 28 • Cory Trimm

2

2

1

Why AI guardrails for public benefits should start at the agency level

Agencies can and should write their own guardrails now, and revise them as AI changes.

Oct 9 • Kara English

4

What AI gets wrong about SNAP, and what we tried to fix it

Part 2 of our series evaluating LLM performance on SNAP policy questions

Oct 8 • Ryan Hansz and Cory Trimm

4

1

September 2026

How we cut agentic AI costs by 44–85% using Vercel’s Eve

Testing a pre-built harness on real California benefits applications

Sep 28 • Kaylyn Van Norstrand

4

How accurately do AI models answer questions about SNAP?

We tested frontier and open-source AI models against 500 caseworker questions

Sep 24 • Ryan Hansz and Cory Trimm

2

6

2

August 2026

How Nava Labs implements LLM guardrails: 5 tactics

Best practices for production applications

Aug 28 • Cory Trimm

1

2

8 multi-agent AI patterns for government

For public service delivery, multi-agent systems must consider privacy, records retention, accessibility, and more.

Aug 7 • Cory Trimm

1

3 gaps in states’ AI governance for public benefits

AI adoption is outpacing governance and evaluation. We should build a feedback loop.

Aug 6 • Kara English

4

Looking at local AI models: experiments with Ollama and n8n

Cost and data security are major concerns for government AI. Local language models can help address them.

Aug 3 • Cory Trimm

July 2026

How AI-generated test docs can improve safety net applications

Synthetic documents, real impact

Jul 15 • Ryan Hansz

2

1

June 2026

How might voice AI reduce administrative burden for caseworkers?

Exploring ways voice to text can support caseworkers who connect people to SNAP, Medicaid and other public benefits.

Jun 18 • Ryan Hansz

1

Say hello to Nava Labs Field Notes!

Join us as we show you under the hood of our experimental work as it’s in progress.

Jun 9 • Greg Jordan-Detamore

2