Research · published 2026-07-28 · 6 min read

We Screened Every Meat Plant in America in One Afternoon

By Mireye Research

Interactive map: all 5,796 analyzed US meat, poultry and egg plants colored by cold-chain fragility

All 5,796 analyzed plants, colored by fragility percentile. Brighter means more fragile. Hover any dot for the plant, or open the chart full screen.

Issue 01 of mireye's research series. Every issue: one industry, one public address list, one question, screened nationally. Data, methods, and code are public, and every value is cited to its federal source.

TL;DR

We screened all 7,185 federally-inspected US meat, poultry and egg plants for cold-chain fragility (heat, grid remoteness, flood) in one afternoon, from nothing but their street addresses.
America's cold chain is weakest exactly where it's hottest: Louisiana, Mississippi, Florida, Texas.
The fragile tail isn't Big Meat. It's small Gulf-coast fish houses, and the smallest plants are six times more likely than the large ones to sit far from the grid.
Humid heat, not dry heat, is what kills refrigeration. Rankings by hot days versus design wet-bulb overlap at a rank correlation of just 0.37.
An AI agent ran the whole study through mireye: 190 addresses profiled per minute, 14,400 API calls with zero server errors, about four dollars of data cost, every value cited.

Build your own

Treat this study as a skill, not a one-off. The recipe:

Take any address list. Suppliers, insurance books, fleet depots, school districts, portfolios. America is full of public ones.
Feed it to mireye. One call per address returns hundreds of cited fields from the catalog. Or skip code entirely and hand Claude the list over MCP, which is how this study was actually run.
Define what fragile means for your world, and let the ranking find your tail.

Everything from this issue is at github.com/Mireye-Labs/cold-chain-fragility: the 5,796-plant dataset, the index, the caveats, the pipeline.

The question

Monday morning we shipped street addresses into mireye-earth. Until then the API spoke coordinates only, and nobody holds a lat/lng. By the afternoon I wanted a real test, not a demo. So: there are 7,185 federally-inspected meat, poultry and egg plants in the United States, and every one of them is a building that dies without refrigeration. How safe is the cold chain at each?

Nobody had answered this, because the pieces live in silos. The FSIS directory is complete by law but carries no physics: names, addresses, and the agency's own reference coordinates, nothing else. The physics lives at NOAA, NREL, EIA, FEMA and EPA. Five agencies, five data models. Joining them used to be a weeks-long GIS project, so a national screen never got scoped: you sampled 200 plants and the tail, where all the fragility lives, never got looked at.

An agent ran it, fast

We pointed Claude at the question over mireye's MCP. It pulled the government's list, resolved all 7,185 addresses, and profiled them at a sustained 190 plants per minute: three fully-cited site profiles every second, sixteen fields each, straight from bare street addresses. About 14,400 API calls with zero server errors, and roughly four dollars of marginal data cost. The whole national study fit inside one afternoon.

None of it works without the address layer underneath. mireye's geocoding engine resolves a bare street address to a coordinate and tells you exactly how it earned it: every match carries an accuracy tier (rooftop, nearest rooftop, or street interpolation), and an address too vague to trust, a bare town or a ZIP, is refused outright instead of silently guessed. That honesty is measurable: across all 7,185 plants, our coordinates landed a median of 18 meters from the government's own reference points. And the engine is growing its own front door: /v1/resolve, which takes the same address one layer deeper to return the coordinate plus county, tract, and parcel in a single call.

Then it attacked its own results: every coordinate checked against two independent sources, the ranking re-run under six weightings, every named plant web-verified as real, operating, and refrigeration-dependent. That pass killed our two best headlines (the original #1 was a geocode artifact, 8 kilometers from the real plant). Everything that died in verification is not in this post. The 5,796 plants in the charts are the analyzable core of the 7,185: co-located inspection grants deduplicated, certification-only offices screened out, and non-CONUS plants dropped where the federal heat grids end. Nothing was excluded for failing.

What survived

America's cold chain is weakest exactly where it's hottest. Fragility runs Louisiana, Mississippi, Florida, Texas. That's the bad kind of correlation: heat waves stress the grid at precisely the moment every condenser in the region is working hardest. Same place, same week.

The fragile tail isn't Big Meat. The most exposed plant in the country is a Louisiana seafood operation: no high-voltage substation within ten kilometers, no transmission line within two, inside FEMA's coastal wave zone, facing 85 days of Gulf heat a year. The big processors bought land next to substations; the fish houses sit where the fish are. And the gradient is regressive: the smallest plants are six times more likely than the large ones to have no substation within ten kilometers.

Bar chart: share of each group ranking in the national top decile of cold-chain fragility

The fragile tail is regressive. Seafood plants land in the national top decile at three times the base rate, and exposure falls as plants get bigger. Full screen.

The numbers: 161 plants have no high-voltage substation within ten kilometers. A quarter have no transmission line within two. 337 sit inside FEMA flood hazard areas. Sixteen coordinate-verified plants face all three hazards at once. A single circle of 100-kilometer radius, centered in Waller County, Texas, contains twelve of the hundred most fragile plants in the country. And exactly one of the 5,796 plants we analyzed runs its freezers inside FEMA's coastal wave zone, the zone built for breaking waves. It is also the most fragile plant on our list, and not because of the flood term: flood is a tenth of the index, and on heat and grid alone that plant still sits in the top five of 5,796.

And my favorite finding: hot is not what kills refrigeration. Humid is. Rank the plants by days over 90°F, then by design wet-bulb, the number that actually sizes a condenser. The two rankings overlap at a Spearman rank correlation of just 0.37. A bone-dry Arizona summer at 164 hot days is easier on a freezer than a Louisiana bayou at 85. Almost nobody screens on wet-bulb, because almost nobody has it as a queryable field. We do.

Interactive twin maps: plants ranked by days over 90F versus ranked by design wet-bulb

Same plants, two definitions of hot. Left: days over 90°F, where Arizona and Texas burn. Right: design wet-bulb, where the Gulf burns. Full screen.

History already ran this experiment

Hurricane Ida (2021) tore through exactly the geography this list points at: $305 million in Louisiana fisheries infrastructure damage, cold storage "almost non-existent," villages off the grid for a month. Winter Storm Uri (2021) hit our large-plant screen by name, suspending Sanderson Farms processing across three states. The index did not predict these events; built purely from physical exposure, it lands on the same plants, parishes, and failure mechanism the last five years already demonstrated. Corroboration you can check.

Honest boundaries

We measured distance to grid infrastructure, not utility reliability, and we know nothing about anyone's backup generators. This says nothing about food safety anywhere. FEMA flood zones flip at building scale (twenty meters of coordinate error can change the answer), so no plant got a flood claim unless two independent coordinate sources agreed. A screen tells you where to look, not what you'll find when you get there.

The part that generalizes

One aside the data whispered: the most fragile quartile of plants sits where industrial electricity is cheapest, a median rate of $78 per megawatt-hour against $98 for the safest quartile. Cheap power is not resilient power, which is the whole finding in miniature.

The bigger point: data assembly used to be the engagement, weeks of geocoding and layer-wrangling before judgment got a minute. Now the assembly is an API call, and any address list plus one question is a same-day, cited, national answer. The analyst doesn't even have to be a person for the first pass. Ours wasn't.

What should we screen next? And what's the address list you've been sitting on?

Frequently asked questions

What is cold-chain fragility?

Cold-chain fragility is a facility's physical exposure to losing refrigeration: how much refrigeration-hostile heat it faces (especially humid heat, measured by design wet-bulb), how far it sits from redundant grid infrastructure, and whether it is in a flood zone. This study scored all 5,796 analyzable federally-inspected US meat, poultry and egg plants on those factors, using only cited federal data.

Which US meat plants are most exposed to heat and grid risk?

The most exposed plants are small Gulf-coast seafood and catfish operations in Louisiana, Mississippi, Florida and Texas, not the large national processors. 161 plants have no high-voltage substation within 10 kilometers, about a quarter have no transmission line within 2 kilometers, and 337 sit inside FEMA Special Flood Hazard Areas.

Is this a food-safety assessment?

No. This is a screening-level physical exposure analysis. It measures distance to grid infrastructure, heat, and flood zone status. It contains no outage or reliability data, no information about backup generators, and it says nothing about food safety or regulatory standing at any establishment.

How can I run this screen on my own facilities?

The full pipeline is open source at github.com/Mireye-Labs/cold-chain-fragility. Point it at any CSV of addresses with a mireye API key, choose fields from the catalog at docs.mireye.ai, and a national screen returns the same day. The API also speaks MCP, so an AI agent like Claude can run the entire study from your address list on request.