Methodology
Our approach to classifying circular economy laws
Atlas Circular is a circular economy legislation tracker and analysis engine. The Atlas holds 2,544 measures across 37 regions worldwide — from Australia to Austria, China to Chile, and Kentucky to Kenya. Each one is screened against a consistent set of circular economy criteria mapped to the technical and biological cascades — 9 legal instruments and 33 material and product streams.
A live snapshot — the engine re-runs as bills move and new sessions open. Of the 2,544 tracked measures, 1,064 are distinct enacted laws in force — the figure the homepage globe shades jurisdictions by. The rest are pending, failed, or superseded bills we keep on the record, plus the acts that only amend a law already counted here: those are tracked and searchable like any other, but counting an amendment beside the law it edits would report one regime as two, and would do it hardest to the jurisdictions that legislate by amendment.
What we screen for
- · Extended Producer Responsibility (EPR) — incl. shared-responsibility & reverse-logistics regimes abroad
- · Deposit Return / bottle bills
- · Right to Repair
- · Recycled-Content mandates
- · Financial incentives (grants, tax credits, procurement)
- · Disposal & landfill bans
- · Organics diversion / composting
- · Labeling & Disclosure
- · Preemption (tracked as a countervailing signal)
plastic packaging, paper packaging, glass, metals, electronics, batteries, nickel-cadmium batteries, vehicles, auto switches, construction, furniture, paint, carpet, mattresses, tires, textiles, lighting, thermostats, pharmaceuticals, medical sharps, solar panels, used oil, hazardous materials, mercury, pesticides, microplastics, marine debris, critical minerals, organics, bio-based materials, agriculture, water, biodiversity.
How a bill gets classified
- 1. Ingest. Every bill from all 50 states and D.C. is pulled from Open States and refreshed as it moves, alongside circular-economy law from the EU and national governments worldwide, drawn from each jurisdiction's official source.
- 2. Pre-screen. A curated, weighted circular-economy lexicon narrows the full legislative universe to plausible candidates, so the deeper analysis is spent only on bills that might be relevant. (The specific terms and weights are proprietary.)
- 3. Classify. Each candidate is evaluated against the fixed criteria above and either flagged relevant — with a confidence score, policy instrument, and material tags — or set aside.
- 4. Extract. Relevant bills have their compliance specifics pulled from the bill text: deadlines, covered products, producer obligations, fees, and preemption signals.
- 5. Review. A growing subset is spot-checked by a human, which flips the bill's reviewed marker.
- 6. Re-screen. As a bill advances or its text changes, it's re-evaluated so the record stays current.
Beyond the flag — what we build on it
Classification is the foundation, not the finish. Once a bill is flagged relevant, the same engine turns it into working intelligence across the platform:
- Compliance extraction — deadlines, covered products, producer obligations, fees, and penalties pulled from the bill text, on each bill's deadline & obligation view.
- Structured dimensions — eco-modulation, recycled-content and collection targets captured as comparable, citation-backed envelopes.
- Full-text search — the ingested statute text is indexed, so a search reaches inside the law, not just its title.
- Jurisdiction profiles — bill activity rolled up per state and country on the Leaderboard, with cross-border comparison in Insights.
- Litigation signals — federal preemption and related cases tracked alongside the bills they touch on Federal Actions.
- Research — natural-language questions answered against the corpus with cited sources in Ask the Atlas.
Every layer above is traceable to the classified bill and its primary source — the same auditable standard as the relevance call itself.
Auto-classified vs. reviewed
Each bill is first auto-classified: a language model reads the title, summary, and text and decides whether it touches one of the tracked instruments, with a confidence score and the material streams it affects. Compliance details (deadlines, covered products, producer obligations) are then extracted from the bill text.
A bill marked reviewed has additionally been spot-checked by a human. Anything not yet reviewed carries only the automated call — shown on each bill so you always know which is which.
Classifications are automated and can contain errors; always verify against the primary source before acting. We continuously expand the reviewed set.
Fair questions we'd ask too
These are the things people want to know before they put a card in, so here they are with the answers up front.
Can a team this size really cover 37 regions?
It's a mix: public legislative APIs where a jurisdiction publishes one, and filings collected individually where it doesn't. That's why the corpus is 2,544 measures rather than a claim of totality — and why every one of them carries a link back to its source document. Check any of them against the record before you trust the rest.
Is this a language model making things up?
The database computes; the model only writes. Counts, dates and aggregates come out of the database as exact queries. The model summarises passages from measures the search has already matched, and it can't report that nothing exists when the matched set isn't empty. Every claim it makes deep-links to the filing it came from.
What happens when the founding rate ends?
Nothing, for you. The founding rate stays with your seat for as long as you keep it, whatever the list price does afterwards. Any change comes with notice first, and your saved research, alerts and exports come with you if you leave.
What if the jurisdiction I need isn't in there?
Tell us and it goes in the queue. Coverage went from US-only to 37 regions in a matter of months, so the constraint is priority rather than capability — and paying seats set the priority. If the gap is a dealbreaker, say so before you subscribe and we'll tell you honestly how long it would take.
See a miscall? Help us correct it.
Flag it →