About Paper Trail | Minerals
Paper Trail | Minerals tests whether trade and corporate-registry data can detect Chinese battery materials being routed through factories in other countries before they reach US battery supply chains. It traces a dozen cases, from graphite anode plants to battery-cell and nickel plants. It uses the Sayari knowledge graph through Sayari's MCP server.
Purpose
Since July 2025, US battery makers can lose clean-energy tax credits if too much of a component's cost comes from a prohibited foreign entity. In March 2026 the US International Trade Commission declined to impose duties on Chinese anode material. Together these shift the compliance question from where did the goods ship from? to who owns the plant they came from? Answering that takes ownership data and trade data together.
It contributes five things:
- Traced supply chains, one case each, each answering one question with the evidence behind every link (Case Files, Explore).
- Practical checks for buyers: who supplies a given US company (Check a Buyer), plus ownership screening and a watch list (Due Diligence).
- A reusable screening playbook: a Sayari MCP call sequence and red-flag rubric that can be rerun on any battery supplier (playbook).
- An honest scorecard of the tool, showing where Sayari worked, partly worked or failed (scorecard).
How to read it
Every link between two companies has an edge type, shown by both colour and line style, and a confidence badge (how we rate it). Click any receipt such as s017 to see the call that produced it.
Each case's verdict (on Case Files) sorts findings into what Sayari detected (a record shows it), suggested (signals point to it without a record) and couldn't see.
How we rate confidence
Every link on the site gets one of four ratings. The rating describes the evidence, not how suspicious the link is. A Documented shipment can be completely ordinary, and a Lead can still turn out to matter. Each link also shows a one-line "Why" explaining its rating.
| Rating | What it takes | Why this threshold | Example from Case 1 |
|---|---|---|---|
| Documented | A single primary record states the link: a dated customs row naming both companies with a product (HS) code, a company-registry shareholder entry, or the company's own stock-exchange or regulatory filing (e.g. an annual report's subsidiary list). | The record itself is the evidence: anyone can open the same record or filing and check it. A filing counts because the company is legally accountable for it. | BTR Indonesia → Ultium Cells Warren: 3 shipment rows of artificial graphite, Jun–Aug 2025 (s017). |
| Cross-checked | No single record states it in full, but two independent sources agree. One of them can be the company's own statement (a press release or filing) if a second source backs it. | Companies rarely lie about owning their own factories, but a claim from one source can be out of date or wrong. A second source guards against that. | BTR owns the Indonesian plant: BTR's press release plus a trade directory, consistent with the trade link and name. Sayari's registry has no record (s009, t012). |
| Inferred | Our judgment from matching details, such as the same street address, a variant of the same name, or shared buyers. No source states the link. | Matching details make a link likely but not certain. Two companies can share a building, and names get truncated. So we show the reasoning and let the reader judge. | Five Sayari records are one plant: four share Jl. Indraprasta No. 8, Kendal (s006, s022, s023). |
| Lead | One weak signal, usually a name match alone or a link we saw but didn't pull records for. | Shown for transparency so others can follow up. It should never be relied on by itself. | The Jakarta Barat gazette record has the exact name but no address or identifiers (s007). |
Rules we apply:
- Ratings only go up with more independent evidence. A second Sayari tool restating the same underlying record doesn't count twice.
- A Sayari risk flag on its own never raises a rating; flags describe neighbours.
- Each case also has an overall rating: the rating of the weakest link the case's main conclusion depends on.
- The edge type (shipment, ownership, inferred) says what kind of link it is; the confidence says how well we know it. They're separate on purpose.
How to read Sayari risk flags
*_adjacent: the entity is one hop from a flagged entity, through any kind of relationship. It isn't flagged itself.psa_*: the path runs through a "Possibly the Same As" link, a probable but unconfirmed record match.*_origin_subtier: a listed entity appears in tier 2 or 3 of the supplier network in the past 730 days, based on trade data.- We show flags as context and mark risky ones with ⚑. We never present a flag as a finding about the company itself. Definitions come from s003.
Naming companies
We name every company the records show: parents, traders, factories and buyers. The aim is full transparency, so anyone can follow each link back to its source. Being named here means a company appears in trade or registry data. It doesn't mean it broke any rule. US battery makers buying from a Chinese-owned plant abroad may be fully compliant; the PFE rules turn on cost shares and legal tests we don't assess.
Individuals (directors, officers, managers) are named only when they add to the supply-chain picture, for example when the same executive runs a plant and its holding company. Their role is always given, and each name links to the registry record it came from. We don't name nominee or registered-agent directors.
Data sources
Publishing these Sayari-derived summaries was approved by the course instructor (Oct 2026). Sayari data is licensed. The site publishes our own summaries (names, link types, counts, dates, flags) and links to Sayari Graph for verification, never raw record dumps. No API keys live in the browser: all queries were run through MCP in VS Code, and the site only reads curated JSON in /data.
Open-source credits
- Cytoscape.js 3.30.2 (MIT): network graph.
- cytoscape-dagre 2.5.0 and dagre 0.8.5 (MIT): left-to-right layout.
- Fuse.js 7.0.0 (Apache-2.0): call-log search.
- Fonts via Google Fonts (SIL OFL): Source Serif 4, IBM Plex Sans, IBM Plex Mono.
- Planned for Iteration 2: D3 and d3-sankey (corridors), Scrollama (scrollytelling), OpenSanctions and UN Comtrade cross-checks.
Limitations and ethics
- This is not an accusation tool. We write "pattern consistent with" or "warrants due diligence", never that a company is evading anything.
- Most third-country processing of Chinese inputs is legal and commercially normal. Companies are named for transparency (see above), not as a finding of wrongdoing.
- Trade data has gaps: some countries aren't covered, product descriptions are vague, dates lag, and HS 3801.10 also covers non-battery graphite.
- Trade records can't show which input went into which export, so the Chinese-origin share of any shipment is unknown.
- Entity resolution can merge or split companies wrongly. Every "possibly the same as" link is labelled Inferred, with the reasoning shown.
- Whether any company counts as a prohibited foreign entity is a legal test that Treasury hasn't fully defined. We don't make that call.
- Policy is moving fast. Dates are as of 5 Nov 2026. China's suspension of its Oct 2025 anode controls legally ends on 10 Nov 2026. The wider truce runs to 10 Jan 2027, but China has published no notice moving the suspension. A separate pause on tighter graphite checks for the US ends on 27 Nov 2026.
- AI use. The investigation and the site were built with Claude Code, which ran the Sayari and Tavily calls. India Clarke set the question and the method and reviews the findings.
Course context
Built for GLOB1-GC 2620 Beyond the Brand: Harnessing Emerging Tech in Ethical Sourcing (NYU Center for Global Affairs, Fall 2026). The assignment: demonstrate Sayari MCP on a problem that traces contractual, transactional and inferred relationships, show it in class, then update it based on feedback.
It is a companion to my earlier project, Battery Passport Audit, which mocked up EU battery passports and marked their gaps. A passport is a self-declaration; trade data is an independent check. The gaps are where due diligence should look.
Changelog
- Iteration 1.6, 5 Nov 2026. Full re-check of every case and policy date, with corrections. Supply-chain graphs with the same four buckets on every case and arrows that run clear of the companies.
- Iteration 1.5, 3 Nov 2026. Renamed Paper Trail | Minerals. The Trace now plays each supply chain as an animated route on a larger map, with a drop-down for each case beside it. On Explore, details sit beside the graph, each case opens fully in view, arrows show direction, clicking traces a whole chain, and risk flags highlight companies without removing the rest of the chain.
- Iteration 1.4, 18 Oct 2026. Four new cases: Ford and BTR, CATL/Ampace/Linergy (now TDK-owned), CATL Debrecen as an EU comparison, and Huayou nickel. Ownership percentages filled in from company filings, with the remaining unknowns and the documents that would settle them. Wider supplier scans for Ford, Panasonic and LG Energy Solution. "Last checked" dates and a monthly automated re-check. Executive summary, guided tour, link previews and an accessibility pass.
- Iteration 1.3, 18 Oct 2026. Every headline claim re-checked against Sayari, with corrections: Case 1 now counts about 55 shipment rows (some loaded in Singapore); Case 2 adds BTR Morocco's first graphite, from BTR Indonesia (Aug 2026); Case 7's "Jamaica" rows are explained as a shipping hub. New pages: Check a Buyer and Due Diligence. New home page map and findings list. The Explore network is laid out in four columns (owners, suppliers, plant, buyers). Policy timeline updated.
- Iteration 1.2, 14 Oct 2026 (before class). Three more cases (EVE Energy Malaysia, Gotion US, Syrah Vidalia), making eight. A clearer network graph: duplicate records merged, colour by role, hover focus, zoom controls, a details drawer. Each case now opens with its question, answer and why it matters. Relevant individuals are named with their roles.
- Iteration 1.1, 14 Oct 2026 (before class). Expanded from one plant to five cases. New plain-language home page with key numbers, a rerouting diagram and a glossary. "Corroborated" renamed to "Cross-checked", with a full explanation of the confidence ratings and a "Why" line on every link. All companies named.
- Iteration 1, 14 Oct 2026. Case 1 (BTR Indonesia) traced with 24 Sayari calls. Policy timeline verified with 14 Tavily calls; three dates or rates corrected. Pages: The Trace, Explore, Case Files, Method, About. Case 2 (Morocco) logged as a lead.
- Iteration 2: after class feedback.