Automated HS Code Accuracy: What Does GingerControl's 99.86% Measure?

What changed: Clarified the September 2025 ATLAS benchmark, its test set and limits, removed unsupported accuracy comparisons, and corrected the relevant penalty and documentation claims.

The short answer

Automated HS code accuracy depends on the digit level, answer standard and products tested. GingerControl's API uses a Quick classification line that scored 99.86% at the 10-digit level on GingerControl's own internal test of 1,000+ products, a first-party test. In the September 22, 2025 ATLAS paper, GPT-5-Thinking scored 55.5% and Gemini-2.5-Pro-Thinking 31.0% at 6 digits on 200 CROSS-derived test examples. Those historical results use a different dataset and do not provide a direct comparison with GingerControl or a current accuracy ceiling.

Next step:Try the Classification Researcher on one product

In this post
  1. What is a defensible HS code accuracy benchmark for production use?
  2. Why Most Automated HS Code Accuracy Claims Are Meaningless
  3. What GingerControl's 99.86% Measures
  4. The Three Architectural Choices Behind the Accuracy
  5. How GingerControl's Accuracy Compares to Other Methods
  6. What did the September 2025 ATLAS benchmark measure?
  7. The Financial Cost of Sub-90% HS Code Accuracy
  8. How to Verify Automated HS Code Accuracy Before Buying
  9. Frequently Asked Questions
  10. Test the API Against Your Own Catalog
  11. References
  12. Related Articles

What is a defensible HS code accuracy benchmark for production use?

A usable HS code accuracy figure states three things: the digit level (6 digits is the international HS level; U.S. entry reporting generally uses the 10-digit HTSUS code), how the right answer was decided, and which products were tested and by whom. GingerControl's own figure, 99.86% at the 10-digit level, comes from an internal test of 1,000+ products, so hold it to the same three questions. Historical ATLAS results published September 22, 2025 use a different test set and do not rank GingerControl against the tested models.


TL;DR

Evaluate an accuracy claim by its digit level, answer standard and test set. GingerControl's automated HS classification API uses the Quick classification line, which scored 99.86% at 10 digits on GingerControl's own internal test of 1,000+ products, a first-party test. Its design encodes GRI 1-6, checks Section and Chapter Notes, and references CROSS rulings. The ATLAS paper published September 22, 2025 reported 40.0% at 10 digits and 57.5% at 6 digits for its fine-tuned Atlas model on 200 test examples. Different datasets prevent a direct ranking, and that study did not establish whether architecture or model size caused the differences.


Why Most Automated HS Code Accuracy Claims Are Meaningless

If you have ever evaluated an HS classification API, you have probably seen accuracy claims like "95% accurate" or "industry-leading accuracy" without any methodology. Those numbers are usually meaningless for three reasons.

Wrong measurement level. A vendor that reports 95% accuracy at the 2-digit chapter level is reporting on whether the system correctly identified "Chapter 85: Electrical Machinery" versus "Chapter 84: Mechanical Appliances." That comparison hides everything that actually matters: the 6-digit subheading determines the international classification, and the 10-digit HTSUS line determines the actual U.S. duty rate.

Wrong test set. Many vendors evaluate accuracy on curated examples drawn from their own training data or on artificial product descriptions written for benchmarking. Production traffic looks nothing like that. Real product descriptions are partial, ambiguous, and full of supplier shorthand. The accuracy that matters is the accuracy on real production traffic, not on benchmarks.

Wrong ground truth. Some vendors compare their output against other automated systems and report agreement rates as accuracy. That is not accuracy, that is consensus, and consensus among text-matching systems converges on the same wrong answers because they all make the same mistakes.

The honest version of an accuracy claim has three components: the digit level, the ground truth method, and the test set composition.

What GingerControl's 99.86% Measures

GingerControl's figure is 99.86% at the 10-digit level on its own internal test of 1,000+ products, run on the quick classification line its API uses. It is a first-party test, not an independent benchmark. Hold it to the same three questions as any vendor's number: the digit level, how the right answer was decided, and which products were in the test set. The section above explains why each one matters.

The Three Architectural Choices Behind the Accuracy

These three design choices describe GingerControl's approach. Their separate contribution to accuracy was not measured by the ATLAS study.

1. Deterministic GRI 1-6 logic, separated from probabilistic layers

The General Rules of Interpretation are not heuristics. They are the law of HS classification, applied in strict sequence:

  • GRI 1: Classification is determined by the terms of the headings and the relative Section and Chapter Notes
  • GRI 2: Incomplete and unfinished articles; mixtures and combinations
  • GRI 3: Goods classifiable under two or more headings (3(a) most specific, 3(b) essential character, 3(c) last in numerical order)
  • GRI 4: Goods not classifiable under any preceding rule
  • GRI 5: Containers and packing materials
  • GRI 6: Subheading-level classification follows the same rules applied to subheadings of equal level

GingerControl encodes these as deterministic rules in the classification engine. A probabilistic model might "decide" that a composite product is essentially its housing because the description emphasizes appearance, but GRI 3(b) weighs the Explanatory Note factors: the nature of each material or component, its bulk, quantity, weight or value, and its role in relation to the use of the goods. GingerControl records which component gives the product its essential character, and why.

2. Section and Chapter Notes enforced as exclusions, not hints

Section Notes and Chapter Notes are the legal exclusions and inclusions that override text similarity. A nylon monofilament more than 1 mm thick reads like textile yarn, but Note 1 to Section XI excludes monofilament of that size from the textile Section, and it is classified with plastics in Chapter 39. A generic LLM might cite the Note in its reasoning paragraph and still output a textile code. A deterministic enforcement layer cannot do that, because the exclusion is a hard rule.

GingerControl enforces Section and Chapter Notes as hard exclusions. If a Note rules out a heading, that heading is not in the candidate set.

3. CROSS ruling integration during classification, not after

CBP CROSS rulings are precedent for U.S. classification. They are also the closest thing to a working corpus of expert classification reasoning. GingerControl's Classification Researcher reads similar CROSS rulings during the classification and uses them as decision inputs.

The opposite pattern, citing CROSS rulings as decorative footnotes after the code is assigned, is common in marketed AI classifiers. It produces the appearance of legal grounding without any of the legal substance. The distinction matters because precedent informs the actual decision in expert classification, not the rationalization of a decision already made.

How GingerControl's Accuracy Compares to Other Methods

Method Accuracy figure Source / measurement
GingerControl (quick classification line used by the API) 99.86% at the 10-digit level GingerControl's own internal test of 1,000+ products; first-party, not an independent benchmark
Experienced human customs specialist Not measured in a public benchmark No published agreement rate
Fine-tuned LLM (Atlas, LLaMA-3.3-70B) 57.5% at 6 digits (40.0% at 10) Historical ATLAS test, September 22, 2025
Generic LLM (GPT-5-Thinking) 55.5% at 6 digits (25.0% at 10) Historical ATLAS test, September 22, 2025
Generic LLM (Gemini-2.5-Pro-Thinking) 31.0% at 6 digits (13.5% at 10) Historical ATLAS test, September 22, 2025
Keyword / text-matching API No published figure No public benchmark
Database-lookup classification No published figure No public benchmark

These are historical results for the models and 200-example test set in the September 22, 2025 ATLAS paper. GingerControl was not tested; its internal result uses a different dataset. The figures do not establish a current accuracy ceiling or a direct vendor ranking.

What did the September 2025 ATLAS benchmark measure?

The ATLAS paper published September 22, 2025, by researchers at Flexify.AI, evaluated the named model versions on 200 CROSS-derived test examples. Its historical exact-match results included:

  • Fully correct 10-digit classification (its fine-tuned Atlas model): 40%
  • 6-digit classification (its fine-tuned Atlas model): 57.5%
  • Comparison: The fine-tuned Atlas model (LLaMA-3.3-70B) outperformed GPT-5-Thinking by 15 percentage points and Gemini-2.5-Pro-Thinking by 26.5 percentage points at 10 digits in this historical test

These results describe the tested model versions on that dataset. Commercial classification APIs and GingerControl were not tested, and the study did not define a compliance-grade accuracy threshold.

Atlas exceeded GPT-5-Thinking by 15.0 percentage points and Gemini-2.5-Pro-Thinking by 26.5 points at 10 digits in this test. The comparison did not isolate model size, training, retrieval or legal-reasoning architecture as the cause.

The Financial Cost of Sub-90% HS Code Accuracy

Accuracy is not an abstract metric. Every percentage point of misclassification translates to real dollars across duty exposure, penalty exposure, and audit cost.

Duty exposure. A misclassification can move the base rate and switch Section 301 (including the forced-labor duties), Section 232 or other Chapter 99 layers on or off. On a $1 million shipment, every 10 points of duty is $100,000 over- or underpaid.

Penalty exposure under 19 U.S.C. 1592(c). For violations affecting revenue, negligence is capped at the lesser of domestic value or 2x the lawful duties, taxes and fees lost or potentially lost; gross negligence at the lesser of domestic value or 4x that amount. Where their assessment is not affected, the caps are 20% and 40% of dutiable value respectively. Fraud is capped at domestic value. CBP completed 465 regulatory audits in FY 2025 and collected $235.47 million as a result of importer audits (CBP trade statistics).

Reasonable care defense. Documented methodology matters. The CBP Reasonable Care publication treats consulting a customs expert as evidence of compliance. A classification system that produces the same evidence a customs expert would produce, GRI reasoning chain, Section/Chapter Note analysis, and CROSS ruling references, can support a reasonable care defense.

How to Verify Automated HS Code Accuracy Before Buying

If you are evaluating an automated HS classification tool, ask the vendor four questions:

  1. What digit level is the accuracy measured at? Six digits is the international HS level; U.S. duty depends on all 10. GingerControl's figure is at the 10-digit level.
  2. How was the right answer decided? Ask who set the answer key and how: published CBP rulings, a licensed broker's review, or another system's output. Agreement with another automated system is consensus, not accuracy.
  3. What was the test set, and who ran the test? Its size, how the products were chosen, and whether the vendor ran it or someone independent did. GingerControl's 99.86% comes from its own internal test of 1,000+ products, a first-party test.
  4. Where does the reasoning come from? Without the reasoning behind a code you cannot audit a miss or show reasonable care. At GingerControl the reasoning comes from the Classification Researcher and from GingerOS; the API returns codes and duty layers without a reasoning report.

A vendor that cannot answer all four with specifics is giving you a marketing number, not a compliance one.

Frequently Asked Questions

What is the most accurate automated HS code generator?

GingerControl's Quick classification line scored 99.86% at 10 digits on its own internal test of 1,000+ products, a first-party result. The September 22, 2025 ATLAS paper reported 55.5% at 6 digits for GPT-5-Thinking and 57.5% for its fine-tuned Atlas model on a separate 200-example test set. These figures do not establish which current commercial tool is most accurate. GingerControl's design encodes GRI 1-6 as deterministic legal logic.

What can lower an AI HS classification API's accuracy?

Incomplete product facts, outdated tariff data and misapplied legal Notes can lead to incorrect classifications. Test an API on your own reviewed products and inspect how it handles ambiguous goods. The September 22, 2025 ATLAS study measured code matches for named models on 200 examples; its Atlas result of 57.5% at 6 digits does not explain the performance of current commercial APIs or establish their accuracy ceiling.

How is HS code accuracy measured honestly?

An honest accuracy figure states three things: the digit level (6 digits internationally, 10 for U.S. duty), how the right answer was decided, and which products were tested and by whom. GingerControl's 99.86% at the 10-digit level is its own internal test of 1,000+ products, a first-party test, so ask us those three questions as you would any vendor. Vendors that cannot answer them are usually reporting a marketing number.

Is automated HS classification more accurate than a human specialist?

Consistency and accuracy are separate questions. Evaluate repeated outputs and correct-code matches on the same reviewed products. GingerControl's internal test and the historical ATLAS test do not compare its workflow with licensed brokers on the same dataset; each code still requires review before filing.

Does 99.86% accuracy mean 0.14% of classifications are wrong?

Not exactly. 99.86% at the 10-digit level is the share GingerControl's quick classification line got right on its own internal test of 1,000+ products; your catalog is a different set of products, so test a sample of it. Misses are typically ambiguous products that genuinely require human review, products with too little input data, or products at GRI 3(b) divergence points. In the Classification Researcher an ambiguous product gets a question rather than a guess; the API classifies from the description it is sent.

How does CROSS ruling integration improve accuracy?

CBP CROSS rulings are the closest working corpus of expert HS classification reasoning. When the classification engine reads similar rulings during the classification process, the rulings actively inform the decision rather than being cited as decoration afterward. For products covered by binding rulings, this means classification aligns with established precedent, which both improves accuracy and strengthens reasonable care documentation. GingerControl's Classification Researcher reads CROSS rulings during the classification, not after.

What accuracy should I expect for my product catalog?

Test a sample of your own catalog: GingerControl's 99.86% at the 10-digit level comes from its own internal test of 1,000+ products, and your products are a different set. For catalogs heavy in composite products, sets, or products requiring GRI 3 analysis, expect more products to need review; for catalogs of straightforward single-material, single-function items, expect fewer.


Test the API Against Your Own Catalog

The only meaningful accuracy test is your own catalog. If you are evaluating automated HS classification accuracy, classify a sample of your real products and compare against expert-reviewed ground truth.

The Classification + Tariff Calculation API is for enterprise customers and starts with a call; after it, GingerControl issues one key sized to your calling pattern. Book a call.

For a company-wide system rather than an API, GingerControl builds GingerOS, the AI trade compliance platform that runs on your process and the data you have, ERP or not. It starts with a Prototype Week, the step before a quote. See how a project runs.


References

  1. ATLAS: Benchmarking and Adapting LLMs for Global Trade via HTS Classification, arXiv
    Data cited: Historical 200-example test: fine-tuned Atlas (LLaMA-3.3-70B) 40.0% at 10 digits and 57.5% at 6 digits; GPT-5-Thinking 25.0%/55.5%; Gemini-2.5-Pro-Thinking 13.5%/31.0%; no GingerControl comparison or causal architecture test
    Source: arXiv 2509.18400
    Published: September 22, 2025
  2. CBP Customs Rulings Online Search System (CROSS)
    Data cited: CBP precedent rulings used as classification reference
    Source: CROSS Rulings Database
  3. CBP Informed Compliance Publication, Reasonable Care (revised September 2017)
    Data cited: Reasonable care standard, consulting a customs expert as evidence
    Source: CBP Reasonable Care Publication
    Published: September 2017
  4. CBP Trade Statistics, FY 2025 audits
    Data cited: 465 regulatory audits completed and $235.47 million collected as a result of importer audits in FY 2025 (updated October 2, 2026)
    Source: CBP Trade Statistics
    Published: 2026
  5. 19 U.S.C. 1592, Customs Penalties for Negligence, Gross Negligence, and Fraud
    Data cited: Penalty calculation structure
    Source: 19 U.S.C. 1592
Your next step

Shortlisting AI for trade compliance?

Hand this to GingerOS.

Try the Classification Researcher on one product See GingerOS run on your own SKUs

Pick a 30-minute slot on the next page with Chen Cui, our CEO. Bring one notice or one SKU list (an ERP export, a spreadsheet or an email will do) and see what GingerOS does with it.

Reviewed by Michael Weick, LCB / CCS
Customs compliance manager with 42 years of experience (ex Subaru of America, Merck, and Motorola).

More in AI & Technology

Our Risk-Free Promise

Full refund.

If you’re not satisfied before MVP acceptance, get a full refund of your deposit.

Click Me!

See if you qualify