The platform
This is the page for the question a careful buyer asks before handing regulated documents to a small company: is any of this real. Everything below is either a mechanism that is in the product today, or a sentence that says on its face that it is not. The few numbers here are counted at page load from the same files the checks run against, and the ones that do not flatter us are the reason to trust the rest.
No card, no account, and no call to book. Tell us what you make and where you sell it, and we will come back to you.
The core
Everything you read in the product is a projection of three records and the matching between them. There is no fourth thing, and there is nowhere an answer can come from that is not one of these.
The product keeps a correspondence between three records. Facts: what is true about a product, including what is in it, what it claims and where it is sold. Rules: what anyone with power over it demands, from a regulator, a state, a marketplace, a buyer, a certifier or your own written specification. Evidence: what proves those demands are met, such as a measured value, a document or a written statement. The middle is the matching: given these facts, which of these rules apply, and does this evidence answer them. That middle is ordinary code over stored records with no model call on the path, so the same records give the same answer every time and every line names the rule it came from. What comes out is one of a small set of named states: answered, expiring, missing, waiting on a fact that has not been given, or not yet mapped.
What comes out, in full
The correspondence
The matching is the middle. Given these facts, which of these rules apply, and does this evidence satisfy them. That is one question, answered by one function, over versioned copies of all three records and the instant you are asking about.
Six steps, in this order, every time. Nothing here is a summary of something more complicated happening elsewhere: this is the path, and each step leaves a row you can read afterwards.
A file arrives by forward, upload or request link. It is sniffed for what it actually is rather than trusted by its extension, read whatever its layout and whatever its language, and turned into named values with their measures, their dates and the place on the page each came from. Every value carries how sure the read was, and a value read below the threshold waits in your review queue instead of entering the record.
Placing a document against a kind and against a product is a decision, so it is allowed to fail. A document the classifier cannot place is filed as unrecognised, named on your queue, and left for you. It is never assigned to the nearest plausible kind, and an unreadable page is never treated as an empty one.
A requirement is read out of the source the issuer publishes, never out of a consultancy note or an aggregator, and it is stored with the condition under which it applies, what would satisfy it, what happens if it is not met, a verbatim sentence in the source's own language, and a hash of the page that sentence came from. A machine translation may ride alongside the quote and is never allowed to be the quote.
Which rules apply to this product, whether the evidence clears them, which of two limits is stricter, and what has expired: all of it is ordinary code over stored rows, with no model call anywhere on the path. Models turn mess into structure at the two edges and are never allowed to decide anything in the middle.
Every run records exactly what it read: the fact rows, the evidence rows, the rule pack versions, the engine version and the instant it was answering as of. Ask the same question of the same records tomorrow and the answer is identical, line for line, because it is a function of them rather than an opinion about them.
Every anchored source is enrolled in monitoring with its own cadence and its own maximum tolerated age. A daily job fetches what is due, normalises it the same way every time, hashes it, and compares the stored quote against a fresh fetch. Drift opens a task naming the source and the rules it feeds; drift, an age past budget, or a recorded contradiction each lower a rule a tier the moment the rules load. Nothing rewrites a rule from a diff by itself.
The engine is not written per category. Rules are data, so a market or a category is a set of rule records with their own citations, and opening one is an addition to the library rather than a rebuild. Every rule that arrives faces the same anchoring, the same quote and the same gates as everything already in it, which is why nothing here is claimed before it is mapped.
The split
This is the one design decision the rest of the platform rests on, and it is the reason an answer here can be checked instead of believed.
A check records exactly what it read: the fact rows, the evidence rows, the rule pack versions, the engine version (today 0.4.0), and the instant it was answering as of. What we said on a given day is therefore a function call rather than an excavation, and the same inputs give the same answer on any machine, forever.
A rule's condition can answer three ways: yes, no, or unknown. Unknown travels: the answer names the fact it is waiting on rather than reading a blank as a pass or a failure.
A check can land in nine states. One of them says a fact is missing and names it. Another says this part of the map has not been drawn yet and names the cell. Neither is allowed to roll up as green, which is what keeps a readiness figure from being a decoration.
Two limits stated in measures that do not compare are not turned into each other. The row goes to review with both numbers shown, because a guess that looks like arithmetic is worse than an open question.
Where two authorities bound the same value, the stricter one is applied and the other is recorded beside it rather than discarded. Your own written specification competes as an issuer and is allowed to beat the public limit.
Facts, evidence, answers and decisions are superseded, never overwritten. That is what makes the history worth reading and what makes a replay of an old answer honest.
One file
Read, placed as a kind, values taken with the position each one came from, matched against the rules that select the product, and the field that could not be read named rather than guessed. The last stage is the one the rest of it rests on.
What happens when a document arrives, in five stages. One: it arrives, at an address generated for your account or dropped on the page. Two: it is placed as one of the kinds the classifier knows, and anything it cannot place is filed as unrecognised rather than guessed at. Three: values are read from their positions on the page and stored with those positions and a confidence. Four: the readings are matched against the rules that select this product, and every line comes back with the authority that set it and the citation to read it. Five: anything read with low confidence is named and sent to a person, and the row stays without a reading rather than being filled with a guess.
01
A file lands at the address we generate for your account, or you drop it on the page. There is no form to fill in first.
Forwarding an email is the whole intake step.
02
Before anything is read, the document is placed as one of the kinds the classifier knows. That decides which reader runs and which rules it can ever answer.
Anything the classifier cannot place is filed as unrecognised rather than guessed at.
03
Each value is read from its position on the page and stored with that position and a confidence, so a reading can be opened back to the spot it was taken from.
Example values, invented for this drawing.
04
The readings meet the rules that select this product, and each line comes back with the authority that set it and the citation to read it.
Named here without their limits on purpose: the real limits and citations come from the shipped rule library, which you can read without an account.
05
A field read with low confidence never enters the record as a value. It goes to a person, and the row stays empty until one answers it.
This is the whole difference between a reader you can trust and one you cannot.
The roster
There are more agents in build. They get named here when you can run them, and not before. A roster is a promise, and this buyer can check.
Reads every document on arrival
Reads any document a supplier sends, in whatever layout it arrives, and files it against the right product and lot.
Checks every value against the rules
Checks every extracted value against the requirements resolved for that product and against your own specification, stricter limit applied, each line carrying its citation.
Holds the record and hands it back
Holds every document traceable back to the original bytes, answers questions from the file with citations, and hands the whole thing back whenever you ask.
Asks your partners for what is still owed
Turns every document a partner still owes you into a request that is already written: a no-login upload link and a message you can paste straight into your own thread.
Says what is ready and what is not
The full-stack readiness picture on demand: what is present, what is stale, what is missing, what is unreconciled, and what fixing it costs. This is the free check.
Builds the pack a reviewer asks for
Builds the outbound pack a reviewer accepts, from templates, complete, ordered the way the recipient expects, and pre-flighted before it goes.
Where a card says a leg is still in build, believe the card
Labels
ShippedThe record is real: a label version per product per market, with statuses, parallel drafts, and lots pinned to the version they were printed under. That is the part a recall or a border check asks for.
So is the engine. Each market resolves its own required-element checklist from the library, with the section behind every element, and the artwork is read into evidence the same way a certificate is: low-confidence reads go to review rather than into the record, and an element we could not read comes back unclear rather than present or missing. Generation runs the same rules in reverse. Your profile, your market and your numbers come back as a complete content specification and a clean draft, with every element carrying its citation and every gap naming the record it still needs.
Artwork is not generated. There is no layout, no typeface and no print file: what comes out is the words and the numbers for whoever sets the artwork. Typography, type sizes and where on the package a statement sits are not checked either, and every report says so on its face.
The library
A rule is not typed in once and then trusted because it is old. It is anchored, quoted, attacked, and watched, and each of those is a mechanical step that leaves a record you can read.
The registry today names 36 watched sources from 17 issuers. The fastest is looked at every 1 day, and the tightest maximum age before the rules it feeds start degrading is 7 days. Read at page load from the same registry the daily job runs from.
All six are mechanical and all six are logged. No signature grants anything: not the founder's, not a consultant's, not yours. A human review can be recorded beside a rule as a marker, and it is never an input to any of these.
| Gate | What it asks | Clears it today |
|---|---|---|
| Primary-source anchoring | Is the anchor URL on the issuer's own official domain? | 30 of 57 |
| Quote verification | Does the verbatim quote still appear in an independent re-fetch of the source? | 0 of 57 |
| Cross-language agreement | Do the English-channel and native-channel readings state the same parameters? | 44 of 57 |
| Adversarial refutation | Did the rule survive a pass whose only job was to kill it? | 0 of 57 |
| Structural completeness | Is the rule checkable: closed predicates, real evidence contract, consequence and dates stated? | 0 of 57 |
| Freshness registration | Is the source enrolled in monitoring, with a staleness budget that is not blown? | 17 of 57 |
A gate that could not run is not a pass. It answers “could not be checked”, it says why, and it blocks promotion exactly as a failure would.
Where the library stands today
57 rules carry a receipt. 0 of them clear all six gates, which is the only tier allowed to report a failure. 57 sit one tier below that: shown, cited, and unable to report a failure at all.
Publishing that number is the point of having the gates. They are mechanical, so they hold our own library back exactly as hard as they would hold anyone else's, and the honest reading of a young library is that most of it is still advisory. A rule below the top tier is downgraded in the engine before it ever reaches your product, so the tier is not a label on a page, it is what the code will let the rule do.
Receipts last evaluated 2026-09-02, against the real dataset. Counted at page load from the committed receipt file, which is the same file the checks read.
This is the product's own comparison table, running on this page. Every limit and every citation in it is read from the library above at page load. The measurements beside them are an invented example, labelled as one.
Lot L-0000-EXAMPLE
Example product, one market
Example dataScroll the table sideways for the reading, its citation and its note.
| Analyte | Required | Found | Method | Status |
|---|---|---|---|---|
| Lead (Pb)lead | ≤ 0.5 µg/day | 0.21 µg/dayTitle 27 CCR §25705/§25805Two issuers bound this one. The stricter is applied, the other (10 µg/day) is kept beside it. | ICP-MS | Pass |
| Cadmium (Cd)cadmium | ≤ 4.1 µg/day | 0.80 µg/dayTitle 27 CCR §25705/§25805 | ICP-MS | Pass |
| Arsenic, inorganic (As)arsenic_inorganic | ≤ 10 µg/day | -Title 27 CCR §25705/§25805Not on the example report, so the row is missing rather than passing. | - | Missing |
| Mercury, total (Hg)mercury_total | ≤ 15 µg/day | not detected, detection limit 20 µg/dayUSP <2232>A non-detect clears a limit only when its detection limit sits at or below it. This one sits above, so the row goes to a person. | ICP-MS | Needs review |
| Total Aerobic Microbial Count (TAMC)total_aerobic_microbial_count | ≤ 1000 cfu/g | 240 cfu/gUSP <2023> | USP <2021> | Pass |
| 3 pass · 1 missing · 1 needs review | ||||
You can read the same tables, with every citation, in the requirements library, and the reasoning behind how a market gets mapped in how it works.
Reliability and security
Four commitments that are in the terms rather than only on a page, and four facts about where the data actually sits.
The subprocessor list, the retention windows and the deletion route are all on the security page, named rather than summarised.
The meter
One thing is counted, and it is products you actively manage. Four things are never counted, and they happen to be the four a tool like this is usually sold by.
The whole ladder, the calculator, and the date each figure took effect are on the pricing page. Every price on this site is produced by the same function that charges the card.
Compliant Always organises and checks compliance documents. It is not legal or regulatory advice, and it does not guarantee the outcome of any review by any platform, retailer, or agency. You remain responsible for your products and your submissions.
Every market, all six agents, the whole record and full export, on the products you tell us about. Get in touch and we will walk you through it.
No card, no account, and no call to book. Tell us what you make and where you sell it, and we will come back to you.
