(1) Creating agentic benchmarks for animal welfare (2) Automating the estimation of a company's animal footprint

Creating agentic benchmarks for animal welfare

Why this matters: AI agents are starting to take real operational control of food-production systems, so we need ways to test whether an agent will actually notice and act on animal-welfare problems instead of quietly trading them away for yield or cost.
The core idea: Build a set of simulation benchmarks, one for each type of farm. In each one, the AI agent plays the role of the farm's management software. It runs the farm across a full production cycle and makes a steady stream of ordinary operating decisions. Some of those decisions quietly involve an animal-welfare tradeoff, like ventilation versus energy cost, stocking density, culling policy, or how fast to respond to disease. We score how well the agent handles those welfare decisions.
Every scenario is fully deterministic. It's one fixed, pre-written world that reacts to the agent's actions the same way on every run. That's the main design choice. A fixed world is reproducible, it lets us run many different models against the exact same situation, and it makes the comparison between models fair and cheap. There's no live human running the test and no drift from one run to the next.
More than just the animals: Real farm decisions ripple out to people, so the benchmark scores more than animal welfare alone. It also tracks how the agent's choices affect the farm's stakeholders, meaning the workers on site, the consumers eating the product, and the surrounding community. And it tracks the plain financial profitability of the farm, because welfare and profit often pull against each other and we want to see how the agent balances real tradeoffs rather than optimizing one number in a vacuum. A good agent should protect the animals while still running a viable business and treating the people involved fairly. Moreover we want our benchmarks to be adopted by labs and a wider community than the animal welfare community.
Examples of farms could include:

• Cage-free egg-laying hens
• Broiler chickens
• Pigs
• Dairy and beef cattle
• Shrimp and fish farming
• Insects
• "Futuristic" or speculative systems like cultivated-meat facilities or fully automated farms, to see how agents handle welfare in new setups where the norms aren't settled yet

Expected outputs:

• A reusable framework for writing these deterministic farm-welfare scenarios
• A growing library of scored benchmarks, one per farm type, each with a welfare rubric grounded in real life tradeoffs made by farmers every day (or they could in the future)
• Comparison results showing how different frontier models handle the same welfare tensions, and how they balance animal welfare against profit and the people involved
• Public reports so labs and welfare researchers can track whether agentic models are getting better or worse at protecting animals

Automating the estimation of a company's animal footprint

Companies report a carbon footprint, but almost none report an animal footprint, meaning how many animals (farmed land animals, farmed and wild fish, shrimp) their products account for. Right now animal activists usually work this out by hand, one company at a time.
The aim of this project is to automate that whole process, as much of it as possibly can be. That means every stage, from the research of finding and reading the source material, to the calculation of turning products into a number of animals, to the judgment calls in between about which assumptions to trust when the data is thin. Anywhere a human currently sits and does this manually, we want to see how much of it a system can do on its own.
There are many ways to approach this, and part of the project is figuring out which combination works best. The team will explore different data sources, different ways to estimate what a company actually sells, and different methods for turning that into a number of animals. Weighing those tradeoffs is a real part of the work.
AI is central to this. A lot of the raw information lives in messy, unstructured places like product catalogs, menus, and company reports, and pulling clean data out of that at scale is exactly what modern language models are good at. There's room to use AI across the whole pipeline, from gathering and structuring the data, to filling gaps where information is incomplete, to reasoning through the assumptions the estimate depends on. Part of the work is deciding where AI genuinely helps and where a simpler, more reliable method wins.
A key goal is for the system to be language agnostic. Company data comes in many languages, and the estimate shouldn't depend on the company operating in English. If it works across languages, it can be applied to companies anywhere in the world, not just in one market.
There's also a forward-looking side. Once you can estimate a footprint from current data, you can start projecting how it might change, for example as a company grows, shifts its product mix, or as a whole sector trends one way. That turns the tool from a snapshot into something that helps anticipate where the impact is heading.
Why it matters: NGOs run campaigns against specific companies, and activists have to decide where to focus limited time and money. A footprint estimate tells them which companies account for the most animals, which makes campaigns and funding decisions far better targeted.
Expected outputs:

• A working pipeline that takes a company and returns its estimated animal footprint, with an uncertainty range.
• A clear writeup of the method and assumptions so others can reproduce and trust it.
• Getting this into the hands of NGOs and used in real campaigns.