Recycling-capacity scan

A desk scan for circular-economy practitioners. It answers four questions about one material from one origin:

1. Who can take it (recyclers, and separately the collectors that feed them)?
2. How much can they take, and is that figure current?
3. What condition must it arrive in?
4. How does it get there legally, and what does that cost?

It was built from a test run on nylon fishing nets from Victoria, Seychelles (azvai.com, November 2026). In that test an unchecked AI answer named plausible companies with guessed capacities and prices, and got the shipping route wrong. The method below exists to catch those errors.

It does not replace the phone calls. Gate fees, current appetite, acceptance specs and freight quotes come from the facilities and a forwarder. The scan ends with a call list, not a decision.

Inputs to collect first (ask if missing)

Workflow

Step 1. Candidate discovery (fast, wide, nothing trusted yet)

• Search in the languages of the destination countries, not only English (French, Spanish, Italian, Norwegian and Lithuanian sources were all needed in the nets test).
• Search for the material, for its waste codes and for the projects that already moved it: EU-funded projects, NGO collection schemes, brand take-back schemes, port and EPR schemes.
• Ask who already handles the material at the origin. A local collector usually already has a buyer and a spec, and that is the best lead in the scan.
• Put every candidate in the ledger as pending. A model's own recall is fine for discovery and worthless as evidence.
• If the material is rigid (pipe, crates, floats), search the general regrind market for that polymer as well. Most fishing-gear recyclers take nets and rope only.

Step 2. The claim ledger

Use templates/ledger.csv (header only) and copy the style of templates/ledger-example.csv. One row per claim, not per company, so every number has its own source, date and verdict. Read references/verification-protocol.md before grading.

Rules of thumb:

• Every facility gets an exists row and an operating row. The operating row states the current status in plain words ("Not operating. Liquidation opened 2025-09-02"); the checker drops a facility whose confirmed status says it has closed.
• A capacity value names its basis and scope, for example "15,961 t post-consumer waste collected in 2025, all feedstock, not nets only" (basis words: input, output, nameplate, collected, planned).
• "Nothing published" is a finding: a row with verdict unverifiable, the page you searched as the source, and a note.

The checker gives each facility one status:

Step 3. Verification passes (in this order)

1. Exists and operating. The company register first (Companies House, annuaire-entreprises.data.gouv.fr and BODACC, Brønnøysund, national registers), then the company's own site and its latest dated news. Closed, in liquidation, or silent for more than two years is a stop. Where the regulator publishes permits (SEPA, Environment Agency, national registers), check that the facility may accept this waste.
2. Process and polymers. Mechanical or chemical, which polymers it takes and which it refuses. A recycler of PE/PP nets is not an outlet for nylon, even if old articles say it is.
3. Capacity. Record the basis (input, output, nameplate, collected, planned), the scope (nets only or all feedstock) and the date. A 2011 nameplate figure is history, not capacity.
4. Input spec. Only the facility's own documents or a call count. Note what is missing; it becomes a call question.
5. Terms. Prices and gate fees are almost never published. Mark them unverifiable and move them to the call list. Never let a model's price range into a report as a fact.

Run the checker after each pass:

python scripts/check_ledger.py ledger.csv --as-of YYYY-MM-DD --check-urls --out ledger-report.md

It fails when a claim is "confirmed" by a model answer or an aggregator alone, has no date, gives a capacity without a basis, or cites a page that returns 404 or 410. It warns about sources older than 24 months, pages that block bots or fail to load, and verified facilities whose input spec is still unknown. Fix every ERROR before writing anything. Install certifi (pip install certifi) so --check-urls uses a current certificate bundle; without it some Python builds report every https page as broken.

Step 4. Shipping route

Read references/waste-shipment-eu.md. Classify the material as it will actually leave the origin (after any stripping and sorting), then work out the procedure from the origin's OECD and Basel status. Quote the article numbers. For a UK origin or a move within one country, the reference only points to the regulators; read their current guidance and quote it. Island origins add a sea leg even within one country. Freight rates come from quotes, not from the web.

Step 5. Outputs

1. ledger.csv plus ledger-report.md (the checker's output).
2. A short report: verified outlets first, then collectors, then dropped candidates with the reason. Every figure carries its date and source.
3. call-list.md: per facility, the open questions from references/call-script.md.

Model delegation

Discovery and first-pass verification are read-heavy, so they can run on a cheaper model or in parallel subagents, one batch of facilities each, all writing ledger rows in the same format. The lead model spot-checks every confirmed row on capacity and terms against the quoted source before the report.

Files

• references/verification-protocol.md: evidence grades, liveness checks, capacity traps, the error types seen in testing
• references/waste-shipment-eu.md: EU import and export rules under Regulation (EU) 2024/1157, Basel plastic entries, waste codes
• references/call-script.md: what to ask a recycler, a collector and a forwarder
• templates/ledger.csv: empty ledger with the header
• scripts/check_ledger.py: the ledger checker (standard library only)
• evals/: fixture with one seeded fault per facility and the unit tests (python -m unittest discover -s evals)
• examples/nylon-nets-seychelles-2026-11/: the ledger and report from the original test