- Maker history
- prior lines · 171 maker's past refusals · 10 prior refused filth decomposition · 1
- Import alerts
- country product red listings · 15
- Broker history
- prior lines · 39,133 broker's past refusals · 21
- Country record
- prior lines · 373,303 country's opened share · 0.017 prior refused of examined · 0.186
- Product
- ready to eat · 0.38 raw animal protein · 0.01 seafood · 0.01
Import inspection
Food imports to inspect first
A model trained on 25 years of Food and Drug Administration (FDA) refusals ranks every food line arriving at the border by the chance an inspector finds a problem — and which one.
0 · Results to action
10 food shipments FDA released that the model would have opened
FDA opened 17,048 of 8,914,045 food lines from July to December 2025, about 1 in 520. This list takes every food line FDA did not open in those months, except lines from makers already on an import alert, ranks them by their highest food-safety hazard with the released model (the percentages are the chance FDA finds that hazard if it opens the line), and shows each combination of maker, importer, and product once, as its highest-ranked shipment. The model never trained on those months. The last column shows what FDA recorded afterward. A later record supports the ranking. No record tells you nothing, because an unopened shipment leaves a record only when FDA looks at that maker again.
| # | Shipment | Maker | Filer | Highest hazard if FDA opens the line |
Next hazards | Why the model flags it | FDA record afterward (within 240 days) |
|---|---|---|---|---|---|---|---|
| 1 |
Betel nut, shelled
India · 2025-08-08
|
Sign in to see | Sign in to see |
Chemical contaminant
over 95%
|
Unsafe additive 68%
|
FDA opened 3 lines of this product from India before July 2025 and refused 3, 3 of them for chemical contaminant. | No record |
| 2 |
Palm nut (palm fruit, sugar palm), dried or paste
Nigeria · 2025-10-04
|
Sign in to see | Sign in to see |
Filth
94%
|
Salmonella 2.1%
Chemical contaminant 2.0%
|
FDA had not opened this product from Nigeria before July 2025. The score comes from the maker, the broker, and similar products. | No record |
| 3 |
Potato chips,fried
Mexico · 2025-08-08
|
Sign in to see | Sign in to see |
Unsafe additive
91%
|
Chemical contaminant 2.1%
Allergen 1.7%
|
FDA opened 81 lines of this product from Mexico before July 2025 and refused 79, 74 of them for unsafe additive. |
|
| 4 |
Corn chips, fried
Mexico · 2025-08-08
|
Sign in to see | Sign in to see |
Unsafe additive
91%
|
Chemical contaminant 2.1%
|
FDA opened 326 lines of this product from Mexico before July 2025 and refused 321, 266 of them for unsafe additive. |
|
| 5 |
Potato chips,fried
Mexico · 2025-10-15
|
Sign in to see | Sign in to see |
Unsafe additive
91%
|
Chemical contaminant 2.1%
Salmonella 1.1%
|
FDA opened 81 lines of this product from Mexico before July 2025 and refused 79, 74 of them for unsafe additive. |
|
| 6 |
Potato chips,fried
Mexico · 2025-10-15
|
Sign in to see | Sign in to see |
Unsafe additive
91%
|
Chemical contaminant 2.1%
Salmonella 1.1%
|
FDA opened 81 lines of this product from Mexico before July 2025 and refused 79, 74 of them for unsafe additive. |
|
| 7 |
Corn chips, fried
Mexico · 2025-10-15
|
Sign in to see | Sign in to see |
Unsafe additive
90%
|
Chemical contaminant 3.6%
Salmonella 1.1%
|
FDA opened 326 lines of this product from Mexico before July 2025 and refused 321, 266 of them for unsafe additive. |
|
| 8 |
Snapper (red, gray, malabar, etc.)
Nigeria · 2025-10-21
|
Sign in to see | Sign in to see |
Filth
89%
|
Salmonella 2.1%
Chemical contaminant 2.0%
|
FDA had not opened this product from Nigeria before July 2025. The score comes from the maker, the broker, and similar products. | No record |
| 9 |
Crayfish (crawfish), fresh water
Nigeria · 2025-10-26
|
Sign in to see | Sign in to see |
Filth
89%
|
Salmonella 2.1%
Chemical contaminant 1.4%
|
FDA had not opened this product from Nigeria before July 2025. The score comes from the maker, the broker, and similar products. | No record |
| 10 |
Palm nut (palm fruit, sugar palm), dried or paste
Nigeria · 2025-10-02
|
Sign in to see | Sign in to see |
Filth
89%
|
Chemical contaminant 2.0%
Salmonella 1.2%
|
FDA had not opened this product from Nigeria before July 2025. The score comes from the maker, the broker, and similar products. | No record |
Where FDA did look again, the ranking agrees: 7% of the 500 highest-ranked shipments had a later shipment from the same maker that FDA opened and refused for a food-safety hazard, against 2.3% of all 7,168,784 food lines scored. Food only (FDA industry codes 02 to 46 without colors); cosmetics, kitchenware, and supplements are excluded. Certification, labeling, and process-control refusals do not count as food-safety hazards. Lines from makers already on an import alert are excluded, because FDA holds them automatically.
1 · Data in
The model learns from FDA inspection results
The training data holds 264,552 refusals since 2001 and 48,096 lines that FDA inspectors opened between 2023-03 and 2025-06. FDA refused 19,426 of those lines. Each card below shows the inputs the model reads for one line and the result it learns.
- Maker history
- prior lines · 2,072 maker's past refusals · 17 prior refused filth decomposition · 2
- Import alerts
- country product red listings · 75
- Broker history
- prior lines · 124,694 broker's past refusals · 290
- Country record
- prior lines · 373,303 country's opened share · 0.017 prior refused of examined · 0.186
- Product
- ready to eat · 0.69 raw animal protein · 0.01 seafood · 0.01
2 · Training
Each new tree corrects the errors of the trees before it
Log-loss on lines FDA opened; lower is better. On the training lines it falls to 0.29675. On lines the model never saw, it levels off at 0.54122.
- Product ±0.20
- Country record ±0.13
- Maker history ±0.08
- Broker history ±0.08
- Import alerts ±0.03
The average change in a line's refusal score when this group of inputs takes the values of a typical line, over 3,000 validation lines.
One depth-3 tree fit to the training lines. The deployed model adds 150 trees in each of its 2 stages.
3 · Answer out
Tamarind, dried or paste
Nigeria · 2026-03-29 · 21TYH14
riskier than 98% of lines
- Country record raises the score +0.10
- Broker history raises the score +0.03
- Maker history lowers the score 0.00
- Maker: earlier lines · refused
- 328 · 0
- Broker: earlier lines · refused
- 3,927 · 30
- Country's opened lines refused last year
- 41%
At 10% capacity: inspect. FDA: refused for filth.
-
Filth 94% · 19.2× the usual rateriskier than 100% of lines · the usual rate is 4.9%
-
Labeling 25% · 2.2× the usual rateriskier than 80% of lines · the usual rate is 11.4%
-
Allergen under 1% · 2.1× the usual rateriskier than 82% of lines · the usual rate is 0.4%
-
Process control 4.2% · 0.9× the usual rateriskier than 71% of lines · the usual rate is 4.4%
-
Certification 1.5% · 0.7× the usual rateriskier than 78% of lines · the usual rate is 2.0%
A refusal can cite several hazards, so these percentages do not add up to the refusal chance. Bar length shows lift on a log scale. The list shows the 5 hazards with the highest lift, out of 11 scored.
The same answer as JSON (import-hazard-output/v1)
{
"schema": "import-hazard-output/v1",
"model": "import-hazard-0.1.1-cm.1",
"line": {
"line_key": "603b39b978",
"arrival_date": "2026-03-29",
"product_code": "21TYH14",
"country": "NG",
"product": "Tamarind, dried or paste"
},
"refused_if_opened": {
"p": 0.9518,
"percentile": 98.0,
"raw": 0.9422
},
"hazards": {
"filth_decomposition": {
"p": 0.9428,
"percentile": 100.0,
"lift": 19.2,
"base_rate": 0.04908
},
"labeling": {
"p": 0.2519,
"percentile": 80.2,
"lift": 2.2,
"base_rate": 0.11367
},
"process_control": {
"p": 0.0415,
"percentile": 70.6,
"lift": 0.9,
"base_rate": 0.04379
},
"certification_eligibility": {
"p": 0.0147,
"percentile": 78.0,
"lift": 0.7,
"base_rate": 0.02039
},
"chemical_contaminant": {
"p": 0.0144,
"percentile": 64.9,
"lift": 0.6,
"base_rate": 0.02526
},
"allergen": {
"p": 0.0076,
"percentile": 82.4,
"lift": 2.1,
"base_rate": 0.00371
},
"unsafe_additive": {
"p": 0.007,
"percentile": 58.2,
"lift": 0.1,
"base_rate": 0.0637
},
"pesticide_residue": {
"p": 0.0025,
"percentile": 68.5,
"lift": 0.1,
"base_rate": 0.03075
},
"salmonella": {
"p": 0.002,
"percentile": 61.3,
"lift": 0.1,
"base_rate": 0.02011
},
"listeria": {
"p": 0.0,
"percentile": 50.0,
"lift": 0.0,
"base_rate": 0.00439
},
"vet_drug_residue": {
"p": 0.0,
"percentile": 86.0,
"lift": 0.0,
"base_rate": 0.00309
}
},
"decision": {
"choice": "inspect",
"capacity": 0.1,
"threshold": 0.76437,
"gate_reasons": []
},
"evidence": {
"maker": {
"prior_lines": 328,
"prior_refusals": 0,
"prior_inspections": 0,
"on_import_alert": false
},
"broker": {
"prior_lines": 3927,
"prior_refusals": 30
},
"country_refused_of_examined_last_fy": 0.408,
"history_months": 26.9,
"drivers": [
{
"group": "country record",
"effect": "raises",
"raw_score_change": 0.097
},
{
"group": "broker history",
"effect": "raises",
"raw_score_change": 0.027
},
{
"group": "maker history",
"effect": "lowers",
"raw_score_change": -0.002
}
]
},
"reference": "14,568 examined, matured test lines"
}
The 50 lines the model would open first for filth
Each cell is the model's chance an inspector finds that hazard if the line is opened; shade = how that line ranks against all 8,071 examined 2026 lines, from the 50th percentile up; outlined = what FDA actually found · 43 of these 50 were refused , 40 for filth
How many to open
8,071 lines FDA examined, January–August 2026, months the model never saw · firms already on an import alert left out (their shipments are held automatically) · Filth: 430 refused lines from 247 firms, ranking quality (ROC-AUC) 0.912
Where the model points across every 2026 line
| Country | Food lines | FDA opened | Labeling | Filth | Unsafe additive | Certification | Process control | Pesticide residue | Chemical contaminant | Salmonella | Allergen | Listeria | Vet-drug residue |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| China | 50M | 1 in 17,000 | |||||||||||
| Canada | 5.8M | 1 in 14,000 | 1.9× | 4.5× | 1.9× | 2.1× | 2.4× | 5.6× | |||||
| Mexico | 4M | 1 in 1,200 | 3.5× | 5.4× | 2.3× | 14.1× | 10.4× | 2.7× | 5.9× | 1.6× | |||
| United Kingdom | 1.3M | none sampled | 1.5× | ||||||||||
| Japan | 1.1M | 1 in 5,900 | 1.7× | 2.1× | 4.5× | 4.3× | 4.2× | 21.1× | |||||
| South Korea | 830k | 1 in 2,100 | 1.5× | 1.8× | 6.3× | ||||||||
| Italy | 740k | none sampled | 1.5× | 1.9× | 3.6× | ||||||||
| France | 730k | 1 in 3,600 | 2.4× | ||||||||||
| India | 510k | 1 in 210 | 7.0× | 3.5× | 1.8× | 10.3× | 3.5× | 26.7× | |||||
| Ecuador | 500k | 1 in 2,500 | 1.9× | 4.9× | 2.1× | 2.6× | 2.7× | 6.3× | 6.1× | 4.9× | |||
| United States | 380k | none sampled | 6.3× | 3.2× | 1.8× | 2.1× | |||||||
| Spain | 260k | none sampled | 1.6× | 7.8× | |||||||||
| Thailand | 250k | none sampled | 5.6× | 3.3× | 2.6× | 2.3× | 2.1× | 4.7× | |||||
| Germany | 250k | none sampled | 1.9× | 3.4× |
Predicted from a 0.5% sample of every 2026 food line (349,864 lines) · darker = more of that hazard than the average of these 14 countries; labels from 1.5× · FDA opened is observed; the rest is not validated, because the model only learned from lines FDA chose to open
import-hazard v0.1.1-cm.1 · dataset 2026-09-18.3 · two stages: refused if opened × which hazard