Industries / Energy, water and waste / Environmental services / Hazardous waste disposal
Identifying and classifying
Identifying and classifying is deciding whether a waste is hazardous in law, and which codes it carries, from where it came from, what it contains and how it behaves in lab tests.
7
Milestones achieved, from 1980 to 2019.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1980
Lists plus four tests
On 19 May 1980 EPA published its first rules for deciding whether a waste is hazardous: lists of named wastes, plus four characteristics (ignitability, corrosivity, reactivity and toxicity), each with a definition or a lab test. EPA had planned to rely mainly on characteristics, including cancer risk and harm to plants, but found no test that waste handlers could run routinely, so it used lists and characteristics together.
LessonWhere a hazard cannot be measured cheaply, regulators name the waste by the process that made it instead.
Credit
US Environmental Protection Agency (EPA)
1984
Small generators brought in
EPA's first generator rules covered only sites that made 1,000 kg or more of hazardous waste a month. The Hazardous and Solid Waste Amendments of 1984 made EPA also regulate sites making 100 to 1,000 kg a month, and set lighter conditions for those making less. EPA estimated about 14,000 large and 75,000 small generators in 2011, and 293,000 to 470,000 sites making under 100 kg a month.
LessonStarting with the largest producers covered most of the waste but few of the sites, and Congress judged the exempt waste could still cause harm.
Credit
US Congress; US EPA
1990
A new leach test
On 29 March 1990 EPA replaced its Extraction Procedure (EP) leach test with the Toxicity Characteristic Leaching Procedure (TCLP). A leach test soaks a waste sample in a liquid meant to imitate the water that seeps through a landfill, then measures what the liquid picks up. The rule added 25 organic chemicals to the eight metals and six pesticides the old test covered.
Credit
US EPA, as required by the 1984 amendments
1998
Basel lists A and B
Two lists added to the Basel Convention entered into force on 6 November 1998. Annex VIII (list A) names wastes presumed hazardous, such as metal wastes containing arsenic, cadmium, lead or mercury. Annex IX (list B) names wastes presumed not hazardous. A listed waste can still be shown to be otherwise by checking it against the convention's hazard characteristics.
Credit
the Parties to the Basel Convention, at their fourth meeting (decision IV/9)
2000
A court limits the test
The TCLP imitates one case: toxic waste dumped in a decomposing town landfill holding 95% municipal waste, over groundwater with a drinking-water well 500 feet away. On 21 April 2000 a US appeals court struck down its use for waste from old manufactured gas plants, because EPA had "not produced a shred of evidence" that such waste went to town landfills. EPA wrote the exemption into its rules on 13 March 2002.
LessonA lab test stands for one way of disposal, and holds up only where waste is really disposed of that way.
Credit
US Court of Appeals for the D.C. Circuit, in Association of Battery Recyclers v. EPA
2015
EU hazards follow chemical rules
From 1 June 2015 the EU tied its 15 hazardous properties of waste (HP 1 to HP 15, such as flammable, toxic and carcinogenic) to CLP, the EU regulation for classifying chemicals, and revised its List of Waste. Of the list's 842 entries, 230 are always hazardous and 178 are the hazardous half of a "mirror entry": a pair of codes where the same waste is hazardous or not depending on what it contains, by the European Commission's count. Ecotoxicity (HP 14) was left for a further study.
Credit
European Commission, Regulation (EU) No 1357/2014 and Decision 2014/955/EU
2019
Test for what is likely
Italian prosecutors charged about thirty defendants, including landfill managers, waste firms and testing labs, with illegal waste trafficking. They had given mirror-entry waste non-hazardous codes on the strength of partial chemical analyses and sent it to landfills for non-hazardous waste. On 28 March 2019 the EU Court of Justice ruled that the holder must work out the waste's composition and look for the hazardous substances that may reasonably be in it, and must class the waste as hazardous if that proves impossible in practice.
LessonA partial analysis is not a classification, and testing for everything is not required either.
Credit
Court of Justice of the European Union, Joined Cases C-487/17 to C-489/17, Verlezza and Others
Watch for
Photos assign EU waste codes
Italtel, an Italian IT company, says its phone app lets staff photograph hazardous waste in on-site storage areas. The app identifies the waste, assigns its European waste code (the six-digit code from the EU List of Waste) and fills in the waste register and transport form. Italtel reports that pilots cut the time spent classifying and registering waste by up to 30%. A photo cannot show which chemicals a waste contains, and that decides mirror entries.
Achieved whenan outside party tests the app's codes against an expert's on a published set of wastes, including mirror entries, and reports the error rate.
Claimed by
Italtel, April 2026
Open
01
Grade models on EPA's answers
A public test of at least 300 hazardous waste identification questions taken from EPA's own letters and memos, each paired with EPA's answer and the section of the rules it rests on, with withdrawn or superseded letters removed. Run current language models on it and publish per-question results. Count a model right only if it reaches EPA's answer and cites the right section of 40 CFR Part 261 or 262.
The briefCredit · unclaimed02
A shared test for classifiers
A public set of at least 100 real lab analyses of wastes that fall under EU mirror entries, such as soils, ashes and sludges, each classified separately by two experienced people with reasons for every hazardous property. Then a scoreboard showing how commercial software, open code and language models classify the same inputs.
The briefCredit · unclaimed
Open 01 · the brief
Grade models on EPA's answers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of at least 300 hazardous waste identification questions taken from EPA's own letters and memos, each paired with EPA's answer and the section of the rules it rests on, with withdrawn or superseded letters removed. Run current language models on it and publish per-question results. Count a model right only if it reaches EPA's answer and cites the right section of 40 CFR Part 261 or 262.
Why it is still open
RCRA Online, EPA's public database, indexes thousands of letters, memos and answers in which EPA headquarters interpreted its waste rules. Recent benchmarks cover the wider field: WuYuEval (4,590 multiple-choice questions on solid waste management, July 2026) and WuYu-EnvLE-Bench (2,521 items built from enforcement cases, with a hazardous waste category). Neither is built on a regulator's written answers to identification questions. The nearest regulatory template, NEPAQuAD (July 2024), builds its questions from US environmental impact statements, not waste rules.
Built with what exists
RCRA Online; the electronic Code of Federal Regulations; a model built for long documents and extraction to draft question-and-answer pairs from each letter, such as Muse Ember (see the Note "The jobs agents repeat"); scoring that counts a question passed only when a model answers it right on all 20 tries (see the Note "All twenty times"); two human reviewers.
First step
Download 50 letters on hazardous waste identification from RCRA Online and write each as a question with EPA's answer and the rule it cites.
Done looks like
300 questions, each answer key checked by two reviewers against the letter, with the reviewers agreeing on at least 95%; published scores for at least five models, with every question's result.
Open 02 · the brief
A shared test for classifiers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 100 real lab analyses of wastes that fall under EU mirror entries, such as soils, ashes and sludges, each classified separately by two experienced people with reasons for every hazardous property. Then a scoreboard showing how commercial software, open code and language models classify the same inputs.
Why it is still open
The 2019 court ruling and the Commission's 2018 guidance set the method, but a search found no published comparison of classification tools on the same inputs. Hennebert and colleagues compared five ways of calculating HP 14 on data from 32 waste streams and 88 more, which tested formulas, not tools. In an earlier INERIS study (2013), two service laboratories analyzed the same 32 waste samples and agreed on whether 84% of them were hazardous, a test of laboratories, not software. HazWasteOnline sells classification from lab data under EU and UK guidance and publishes no error rate. Dsposal's waste-code search finds one of the 842 codes from a plain-English description, without checking composition. The open library go-clp classifies chemical products, not waste.
hazwasteonline.comosti.govhazwasteonline.comdsposal.ukgithub.com
Built with what exists
The Commission's 2018 technical guidance and its worked examples; Regulation (EU) No 1357/2014; go-clp's mixture rules as the base of an open classifier; HazWasteOnline and a general AI model as tools under test; an agent that cites the paragraph of guidance behind each step (see the Note "Findings that cite notes"); a spreadsheet.
First step
Ask the authors of Hennebert et al. (2016) for the analyses of their seven mirror-entry streams and classify them by hand with the 2018 guidance.
Done looks like
100 cases with agreed answers, or recorded disagreement between the two experts; each tool's hazardous or non-hazardous call and list of hazardous properties scored against them; the rate of agreement between the two experts reported.