Open research, July 2026
Half the detail dies on the way to the return
AI waste analytics now classify material at the belt in over a hundred categories, and the industry sells that granularity as regulatory reporting. Every UK channel that consumes composition data accepts between 7 and 47 values. We built the crosswalk, measured the loss, and found that the mapping cannot be a function in either direction.
Detail at the belt
5.907
bits of composition detail in the source vocabulary
Worst channel
38.4%
retained by Simpler Recycling
Answerable nowhere
4 of 10
operational questions no channel can answer
EWC codes reached
13 of 47
granularity that exists and is never addressed
Why this matters in October 2026
Digital waste tracking becomes mandatory for permitted and licensed waste receiving sites in England, Wales and Northern Ireland in October 2026, for Scotland in January 2027, and for waste collectors in October 2027. The legal basis is section 58 of the Environment Act 2021. At the same time, packaging producers are reporting under extended producer responsibility and completing recyclability assessments whose outcomes modulate their fees.
Sitting between the material and those returns is a layer of computer vision. Commercial analyzers over conveyors identify material in more than a hundred classes and their vendors state, correctly, that this taxonomy is mapped into regulatory reporting categories. That mapping is where the interesting question lives, and it is not published by anyone. So we built one from open sources and measured it.
What survives each channel
We measure retained detail as the mutual information between the perception class and the value the channel records, as a share of perception-class entropy. A channel that preserved every distinction would score 100%.
| Channel | Targets | Values used | Worst fan-in | Undetermined | Out of scope | Bits lost | Retained |
|---|---|---|---|---|---|---|---|
| EWC / List of Waste | 47 | 13 | 18 | 3 of 60 | 0 of 60 | 2.728 | 53.8% |
| pEPR categories | 10 | 12 | 19 | 7 of 60 | 19 of 60 | 2.829 | 52.1% |
| RAM 2027 groups | 10 | 12 | 22 | 7 of 60 | 22 of 60 | 2.986 | 49.5% |
| Simpler Recycling | 7 | 7 | 23 | 1 of 60 | 0 of 60 | 3.640 | 38.4% |
Read the EWC row carefully. The catalogue offers 47 six-digit codes for municipal and packaging waste and the crosswalk reaches only 13 of them. Granularity that exists in a standard but is never addressed benefits nobody. These figures are also an upper bound, because a class whose value is undetermined is credited with its whole candidate set.
The crosswalk cannot be a function
Collapse is the failure everyone expects: many belt classes share one reported value, up to 23 of them in Simpler Recycling. Two further failures are less obvious and matter more, because conventional mapping quality measures cannot express either.
Underdetermination. For 7 of 60 classes the perception class does not fix the regulatory value at all. A drinks can is aluminium or steel, and RAM 2027 lists cans under both metals. An unqualified plastic item does not say rigid or flexible, which is precisely the split that pEPR requires of large producers in reporting years 2 to 4 and that RAM 2027 uses to cap flexibles at Amber. The classifier has not made this decision. Whoever completes the return makes it, invisibly.
Out of scope. pEPR cannot represent 19 of the 60 classes and RAM 2027 cannot represent 22, because both schemes are packaging-scoped and a great deal of what arrives on a belt is not packaging. That material does not appear in the return at all.
Standard ontology-matching evaluation scores an alignment on precision and recall against a roughly one-to-one reference alignment. None of these three phenomena is expressible that way, which is why the loss has gone unmeasured.
Which operational questions survive
Retained bits are abstract, so we also asked ten concrete questions an operator, a reprocessor or a regulator would want answered. Each needs two groups of belt classes to stay distinguishable. Collapsed means the channel gives them the same value. Undetermined means a class does not fix its own value, so the distinction survives only if whoever completes the return guesses consistently. Partial means one group is not recorded at all.
| Question | EWC | pEPR | RAM 2027 | Simpler |
|---|---|---|---|---|
| Plastic bottles versus plastic trays and tubsDeposit return and PRF sorting both turn on it | collapsed | collapsed | collapsed | collapsed |
| Flexible film versus rigid plasticFlexibles cap at Amber under RAM 2027, rigids can reach Green | collapsed | yes | yes | collapsed |
| Aluminium versus steel packagingpEPR sets a separate base fee per metal | collapsed | yes | yes | collapsed |
| Fibre-based composite versus plain boardLiquid cartons cannot reach Green, plain board can | yes | yes | yes | yes |
| Drinks cans versus food cansDifferent reprocessing routes and deposit return scope | collapsed | undetermined | undetermined | collapsed |
| Expanded polystyrene versus other plasticEPS has no kerbside route in most UK systems | collapsed | partial | partial | collapsed |
| Paper cups versus paper packagingCups need a separate reprocessing route | yes | yes | yes | yes |
| Packaging versus non-packaging, same materialpEPR liability attaches only to packaging | yes | partial | partial | collapsed |
| Crisp packets versus other flexiblesMetallised film is a distinct sortation problem | collapsed | collapsed | collapsed | collapsed |
| Glass bottles versus glass jarsColour and format both affect remelt value | collapsed | collapsed | collapsed | collapsed |
Two of ten questions are answerable in every channel. Four are answerable in none. Three results are worth stating plainly.
- The metal split is the expensive one. pEPR sets separate base fees for aluminium and steel. The List of Waste offers a single code,
15 01 04 metallic packaging, and no aluminium or steel split exists anywhere in the catalogue. So the distinction that determines the fee cannot be carried by the code that the mandatory waste tracking schema requires. - Metallised film fails everywhere. Crisp packets are a distinct sortation problem and a distinct reprocessing problem, and not one of the four channels has a cell for them.
- EWC wins the one nobody expects. It separates packaging in
15 01 xxfrom municipal fractions in20 01 xx, so it can tell a packaging bottle from a reusable food box. pEPR and RAM simply do not record the non-packaging side.
A return that shows its own uncertainty
Waste analytics products emit a regulatory return as a point estimate. If the crosswalk is not a function, a point estimate is the wrong object: the true figure is bounded, not known. So the engine in this repository emits intervals instead, and names the classes responsible for each band.
Put a composition through the pEPR categories and the aluminium line spans a factor of three. Aluminium carries its own base fee.
| pEPR line | Reported interval | Classes driving the band |
|---|---|---|
| plastic - flexible | 819 | none, fully determined |
| plastic - rigid | 780 to 787 | Squeezable tube |
| Aluminium | 167 to 503 | Aerosol, Drink can, Metal bottle cap, Metal lid, Squeezable tube |
| Steel | 34 to 363 | Aerosol, Drink can, Metal bottle cap, Metal lid |
| Paper and card | 122 to 245 | Meal carton, Other carton |
| Fibre-based composite | 113 to 236 | Meal carton, Other carton |
On the same composition, 39.3% of the material cannot be represented in pEPR at all and 45.6% cannot be represented in RAM 2027. The engine reports that mass separately rather than absorbing it into "other", which is what makes a return look complete when it is not. Every line carries the mapping rows responsible, with their justification and confidence, so any figure can be audited back to the regulator's own text. The demonstration composition is built from annotation counts in open litter photography, which are item counts and not belt tonnages, and it is labelled as such throughout.
How it is built, and what would falsify it
Six SKOS concept schemes, 171 concepts, 1,360 triples. Each scheme records the URL it was transcribed from, its publisher and its licence. Targets are transcribed verbatim from the publisher's own page: the pEPR categories and plastic subcategories from PackUK's regulation 7(11) notice, RAM 2027's assessed groups with their published format lists and permitted red, amber and green outcomes, Simpler Recycling's mandatory streams and named dry materials, and the composition-bearing fields of the Defra digital waste tracking receipt. The List of Waste is reused from the published ewc-onto rather than retyped.
Alignments are emitted as SSSOM, one file per channel. Where a mapping is supported by a regulator's own published format list, that entry is carried as the mapping's evidence and its justification is recorded as lexical matching rather than curation. RAM 2027 is the only channel where this is possible at scale: 36 of its 67 mapping rows are grounded in the regulator's own words, against 5 for Simpler Recycling and 0 for both pEPR and EWC, which publish material names only. That asymmetry is a finding about the schemes, not about the method. Ambiguity is never resolved by guessing: an underdetermined class emits one row per candidate at reduced confidence, and the fork stays visible.
Nothing is published unless a release gate passes. It has two parts: SHACL conformance, and a closed-world check that every identifier used in a mapping resolves to a concept that actually exists. An open-world RDF store accepts mappings to concepts that were never defined, and this catches them.
The gate earned its place immediately
It flagged five mappings pointing at a code that did not exist in the built vocabulary. The cause was upstream. Of 990 labelled classes in ewc-onto v2025-07-08, 970 carry an English label and 20 do not, and exactly three of those hold an English string tagged as German, so no English-language query can see them: 20 03 01 mixed municipal waste, which is the most heavily used municipal code in the catalogue, plus 10 03 and 03 02 05. The correction is applied in the build, asserted against the upstream string so it fails loudly if upstream changes, and recorded on the concept. It has been reported rather than absorbed.
What this means for buyers and builders
- For MRF operators procuring analytics: a vendor claim that a taxonomy is "mapped to regulatory reporting categories" is not a quality statement until the mapping is inspectable. Ask which of your questions the mapping can still answer, and ask what the return does with material the scheme cannot represent.
- For producers and compliance schemes: where a figure depends on a distinction the belt cannot make, such as aluminium against steel, the number you file is a choice someone made and not a measurement. It should carry a band.
- For Defra, PackUK and the agencies: the granularity already exists in the List of Waste and is not being reached, while the distinctions that set fees are absent from it entirely. A one-code change, splitting
15 01 04, would make the metal question answerable through the mandatory channel. - Honest boundary: the source vocabulary here is open litter photography, not a commercial MRF belt, and the weighting is item counts rather than tonnages. A larger source vocabulary makes every retention figure worse rather than better, so these are conservative. The curation is expert judgement, declared as such, with a rationale on every row.
Open, and reproducible
Vocabularies, alignments, metrics, the reporting engine, the SHACL shapes and the release gate are public under CC BY 4.0. Every figure on this page regenerates from three commands in the repository, and the entropy figures were additionally recomputed from the published mapping files independently of the build, matching to four decimal places. An interactive explorer ships with it: a single self-contained file with the retention chart, the answerability matrix, the interval chart and a filterable row for every belt class.
Source vocabulary: TACO, 60 annotated classes and 4,784 annotations, MIT licensed. Targets under Open Government Licence v3.0; the List of Waste ontology under CC BY 4.0. A note on data we deliberately did not use: the Environment Agency's Waste Data Interrogator would be the natural tonnage weighting, and it is unusable for a published artefact because it carries the Environment Agency Conditional Licence, which permits internal or personal use only, for one year, without sublicensing. Mass weighting needs an Open Government Licence source instead.
Related work: we are working with WRAP on a food loss and waste data taxonomy for the Food Programme and global Food Pact Network; the industrial and construction crosswalks apply falsifiability measurement to other standards families.
