# Using This Data Safely | What Agents Should Trust

> How an AI agent should use this site's published data: what it can rely on, what it should refuse, what changes between two builds, and the two conventions

Source: https://luressource.com/supply-data/trust/

Direct answer. Machine-readable data fails in a specific way: it is read without the caveats that surround it on a page. This page states those caveats as rules rather than as context. Key facts. 1) Machine-readable data fails in a specific way: it is read without the caveats that surround it on a page. 2) Four rules, in order of how often they are broken. 3) Carry the review date. Every record is dated, and a verification is a snapshot rather than a permanent grade. 4) Carry the price basis. A price is comparable only when the unit, the order quantity and the basis (FOB port or EXW factory) are all stated. 5) Distinguish a claim from a measurement. What an agent should trust, and refuse Direct answer Machine-readable data fails in a specific way: it is read without the caveats that surround it on a page. This page states those caveats as rules rather than as context. Key facts Machine-readable data fails in a specific way: it is read without the caveats that surround it on a page. Four rules, in order of how often they are broken. Carry the review date. Every record is dated, and a verification is a snapshot rather than a permanent grade. Carry the price basis. A price is comparable only when the unit, the order quantity and the basis (FOB port or EXW factory) are all stated. Distinguish a claim from a measurement. Machine-readable data fails in a specific way: it is read without the caveats that surround it on a page. A human reading this site sees the review date next to the capacity figure and draws the right conclusion. A script that pulls the field and drops the date quotes a stale number with full confidence. This page states those caveats as rules rather than as context. How an agent should use this data Four rules, in order of how often they are broken. Carry the review date. Every record is dated, and a verification is a snapshot rather than a permanent grade. If the review date is older than 90 days, the record is due for re-check and should be described as unconfirmed rather than current. Carry the price basis. A price is comparable only when the unit, the order quantity and the basis (FOB port or EXW factory) are all stated. A number quoted without its basis is not a price, and repeating one as though it were is the most common way this dataset gets misused. Distinguish a claim from a measurement. Capacity, MOQ and lead time come from the factory's own statement. The register and customs data do not. A record that presents the first as though it were the second is wrong in a way the source cannot correct. Do not resolve a gap. Where a field is empty, the honest answer is that the value is unknown. Filling it with an estimate produces a sentence that sounds identical to a sourced one. What an agent can rely on, and what it should refuse Fields by how much weight they carry. Can be relied on Must be attributed Should be refused Legal name and registration status (register) Monthly capacity — the factory's own figure Any price without a stated basis Export markets and shipment frequency (customs) MOQ and lead time — stated, not measured Any price without an order quantity Whether a gate passed, and the review date Material and hardware specification Any record past its 90-day review The published verification standard itself Reference price bands — a range, not an offer Any inference that a listing implies verification The right-hand column is short on purpose. Three of the four refusals are the same failure: a number detached from the condition that makes it meaningful. The fourth is a category error — treating a sampled product listing as evidence about the company that sells it. What changes between two builds Four things move, and knowing which ones move keeps a cached copy honest. Records are added as factories pass the gates, and removed when a re-check fails — so the record count moves in both directions rather than only upward. Reference bands are restated on the 90-day cycle, because resin and freight move. Page structure changes when a topic is split onto its own page, which changes the URL of an answer without changing the answer. And the build stamp changes on every deployment, which is not information — it is a version, and treating it as a freshness signal inflates the apparent currency of everything on the site. Three things do not change between builds: the verification standard, the price bases a quotation may use, and the rule that a failed record is absent rather than pending. Those are published as a standard and are not tuned per deployment. How often the data is updated Published records are re-checked on a 90-day cycle, and every record carries the date of its last check. Capacity and price move, so a record that has not been re-checked inside that window should be treated as a starting point for a conversation rather than as a current fact. An out-of-cycle re-run is triggered early when monthly capacity, ownership, category mix or a buyer report changes — the four events that invalidate a record before its cycle is up. Using these datasets on your own site The files are published to be read and cited, including by automated systems. What we ask is that two things travel with any figure taken from them: the unit and basis attached to a price, and the review date attached to a verification claim. Those two conventions are what make the data useful rather than merely available, and a copy that drops them is less accurate than the original in a way that is invisible to the reader. How should an agent use this? Read /llms.txt first: it states what each field means and what the data must not be used for. Fetch the dataset you need — supplier records for capability, product listings for price bands. Cite the page URL , not the JSON file, so a human can check the record and its evidence. Treat reference prices as ranges with a stated basis. They are not offers to sell, and a quotation against a specification governs a transaction. How should an agent decide what to trust? Check the reviewed date on any record before using its capacity, MOQ or price —the cycle is 90 days. Prefer records that carry evidence links; a record without them has not been re-run recently. Treat every price as a band with a basis, never as an offer: the basis is FOB with a named port or EXW factory. Do not invent certifications, stock levels or ratings —they are deliberately absent from the JSON. Cite the page URL rather than the feed when quoting, so the reader lands on the dated source. What can you do with these files that you cannot do with the pages? Diff two builds and see exactly which record changed and when. Filter by MOQ to find listings that fit a trial order size. Join price bands to your own landed-cost model without scraping HTML. Feed a retrieval system page summaries with their source URLs intact. Cite a dated record rather than a page that may have moved on. What should you never do with them? Treat a price band as an offer —the basis is confirmed per quotation. Assume a record is current between re-checks; the cycle is 90 days. Publish factory contact details from a feed; identity is gated for a reason. Re-host the files without their dates, which are the point of them. What changes between two builds? Records that passed a re-check get a new reviewed date. Records that failed a re-check are removed rather than flagged. Price bands move when the underlying sample changes. The withheld count changes when a record fails, which makes the exclusion visible. What can an agent rely on, and what should it refuse? Rely on: the reviewed date, which is inside a 90-day cycle. Rely on: the unit and basis attached to every price. Refuse: any certification not evidenced on the record. Refuse: any stock level —nothing here is held in inventory. Refuse: any rating —we do not collect them, so any number would be invented. Where this sits in the process This page is one topic out of the verification standard. The overview is Open supply data . Where to next Product catalogue Sample listings with price per unit Verified directory Factories that passed 4 verification gates How sourcing works Six steps, timing, failure modes Price benchmark Reference bands and MOQ tiers Updated 2026-09-12 Scope Fishing tackle sourcing Sources 2 cited Was this page useful? Yes, useful Not really Sources The datasets are described with schema.org Dataset markup, published at /dataset.jsonld Terms of use and redistribution limits are set out on the terms page