EMPORION — public-record facts for agents that trade

🌐 English · Русский · Español · Français · 中文

An agent that buys, sells or pays needs facts no model can make up: is this VAT number valid, who is this company, what was the euro rate that day, is this wallet on a sanctions list, and can I prove this document existed yesterday. EMPORION answers each from the public record — the European Commission, GLEIF, the ECB, the U.S. Treasury, the EU, the UK and Bitcoin — and signs every answer together with where it came from, when, and the SHA-256 of what the source returned.

Emporion (ἐμπόριον): the trading post of a Greek city, where traders met and settled.

Since 0.2.0 it also answers the questions that come next: is this company in its national register, is its EORI number valid for customs, how old is the domain the shop runs on, which day is the fifth business day in Bavaria or on TARGET2 — and it gives agents two tools no model has: a web page captured as evidence (anchored in Bitcoin) and a timer, including a dead man's switch, that calls the agent back with a signed event.

Capabilities

Capability	Price	What it answers	Source
emporion.vat.check@v1	$0.002	Is this EU VAT number valid today? Name, address, time of the check — and, when you name your own VAT number as requester, VIES's consultation number: official proof of the check in your name	EU VIES
emporion.lei.lookup@v1	$0.001	Who is this legal entity? By LEI (check digits verified locally) or by name (ranked candidates): legal name, jurisdiction, legal form, status, addresses, renewal, direct and ultimate parents	GLEIF (CC0)
emporion.fx.rate@v1	$0.001	Euro reference rate for any pair on any date since 1999, the ECB day it is from, cross rates through the euro in Decimal, an amount rounded to the currency's minor units	ECB
emporion.sanctions.screen@v1	$0.002	Is this wallet address or name on a public sanctions list? OFAC SDN (with ~510 digital-currency addresses), EU and UK lists (UN designations through them, each hit with its UN number), with each list's date and SHA-256	OFAC · EU · UK
emporion.timestamp.stamp@v1	$0.001	Anchor a SHA-256 digest in Bitcoin through the OpenTimestamps calendars; returns a standard .ots proof	OpenTimestamps
emporion.timestamp.verify@v1	$0.001	Upgrade and verify a proof: confirmed when its result equals the Merkle root of the Bitcoin block header EMPORION holds from the Bitcoin network itself (every header's proof of work checked, the chain agreed by two nodes)	Bitcoin
emporion.web.capture@v1	$0.003	A web page as evidence: final URL, every redirect, the IP, status, key headers, the TLS certificate and the SHA-256 of the body (and the body, up to 512 KB); the record's SHA-256 is anchored in Bitcoin	the page itself · OpenTimestamps
emporion.domain.rdap@v1	$0.001	Who registered this domain and when: created, expires, age in days, registrar, status, nameservers, DNSSEC — or registered: false	IANA bootstrap → the TLD's registry
emporion.company.lookup@v1	$0.001	Is this company real and alive? By number or name: name, status, legal form, founding date, address, VAT number, and the register's own fields	Brønnøysund (NO) · SIRENE/RNE (FR) · ARES (CZ) · PRH (FI) · Companies House (GB, with the operator's key)
emporion.eori.check@v1	$0.002	Is this EORI number valid for EU customs? Up to 10 at once; name and address where the operator consented	European Commission (EOS)
emporion.calendar.check@v1	$0.001	Business days in ~250 countries (and regions) or on a market (TARGET2, NYSE, LSE…): holiday, next/previous business day, date + N business days, days between, a year's holidays	python-holidays (offline)
emporion.timer.set@v1	$0.002	A signed POST to your https callback at a time, or a dead man's switch that fires unless you check in every N seconds	—
emporion.timer.checkin@v1 · .cancel@v1 · .status@v1	free	Keep a dead man's switch alive, cancel a timer, read its state (with the timer's token)	—
emporion.sources@v1	free	Every source's licence and attribution, the sanctions lists held now, the latest ECB day	—

flowchart LR
    A[Agent] -->|invoke + payment| H[AIMarket hub]
    H -->|X-Provider-Signature required| E[EMPORION]
    E --> N{allowlist<br/>per source}
    N --> V[VIES]
    N --> G[GLEIF]
    N --> C[ECB]
    N --> S[OFAC · EU · UK lists<br/>refreshed every 6 h]
    N --> O[OpenTimestamps calendars]
    E --> X[Bitcoin P2P network<br/>headers · proof of work checked]
    E -->|answer + evidence:<br/>url · fetched_at · sha256| H
    H -->|signed receipt| A

Trust, briefly

• Evidence in every answer — the source URL, the time it was fetched and the SHA-256 of the bytes it returned, plus the source's licence and attribution.
• Two signatures — the gateway's signed receipt (oracle-core, Ed25519, hybrid ML-DSA when ORACLE_PQC=1) and the request-bound X-Provider-Signature the hub verifies.
• A source that cannot answer costs nothing — the answer is ok:false, and the hub debits nothing for ok:false: a member state's VAT registry that is down, sanctions lists not loaded or older than 48 h, fewer than two timestamp calendars, a block the header chain does not reach yet.
• Outbound requests are allowlisted per source; redirects are followed by hand and only to that source's own hosts; bodies are size-capped.
• The two capabilities that call a URL the caller names (web capture and timer callbacks) reach public addresses only: http(s) on ports 80/443, every address of the name checked (no loopback, private, link-local, CGNAT or metadata ranges), the connection pinned to the checked address, every redirect checked again, and at most 20 captures of one host per minute.

Examples

H=https://emporion.modelmarket.dev
inv() { curl -s -X POST $H/ai-market/v2/invoke -H 'Content-Type: application/json' -d "$1"; }

inv '{"capability_id":"emporion.vat.check@v1","input":{"vat":"NL004495445B01","requester_vat":"DE811569869"}}'
inv '{"capability_id":"emporion.lei.lookup@v1","input":{"name":"Deutsche Bank"}}'
inv '{"capability_id":"emporion.fx.rate@v1","input":{"base":"USD","quote":"JPY","date":"2026-10-11","amount":"100"}}'
inv '{"capability_id":"emporion.sanctions.screen@v1","input":{"address":"0x…"}}'
inv '{"capability_id":"emporion.timestamp.stamp@v1","input":{"digest":"'"$(sha256sum contract.pdf | cut -d' ' -f1)"'"}}'
inv '{"capability_id":"emporion.web.capture@v1","input":{"url":"https://shop.example/terms"}}'
inv '{"capability_id":"emporion.domain.rdap@v1","input":{"domain":"shop.example"}}'
inv '{"capability_id":"emporion.company.lookup@v1","input":{"country":"FR","number":"552032534"}}'
inv '{"capability_id":"emporion.eori.check@v1","input":{"eori":"CZ00177041"}}'
inv '{"capability_id":"emporion.calendar.check@v1","input":{"market":"XECB","add_business_days":5}}'
inv '{"capability_id":"emporion.timer.set@v1","input":{"callback":"https://agent.example/hook","deadman_seconds":3600,"payload":{"release":"escrow-7"}}}'

Through the hub, the same calls are paid and receipted: POST https://modelmarket.dev/ai-market/v2/invoke with "product_id":"emporion".

VAT: the consultation number

VIES issues a consultation number only to a valid EU VAT number named as requester. It is the requester's proof, kept by the requester: pass your own number as requester_vat and the consultation number in the answer is evidence, in your name, that you checked your counterparty at checked_at. Without it the check is still made; the answer says how to get the number.

FX: which day

A rate "on" a date is the latest ECB publication on or before it, and the answer names that day (ecb_date): a Sunday is priced at Friday's rate, visibly. A currency the ECB stopped quoting (RUB since March 2022) is refused rather than answered with a years-old rate.

Sanctions: a lookup, not a determination

A wallet is matched exactly (EVM and bech32 addresses case-insensitively). A name is matched as a set of words — order, case, accents and legal forms ignored ("DOE, John" is "John Doe"), a word within a typo counts 0.9 and never reads as the same name, and every asked word found in a longer name (a patronymic, a middle name) still matches. The default threshold is 0.85; each hit says which name matched and its score.

A hit is a reason to look closer; a miss is not a clearance. Names collide and aliases are incomplete. Every answer says so, and names the lists it used with their publication dates and SHA-256. An outdated list is refused, not used.

UN designations come through the UK and EU lists. The UN Security Council's own file carries no reuse licence, so EMPORION does not read it. The UK Sanctions List (Open Government Licence v3.0) gives every UN designation the UK implements its UN reference number, as the EU file does for some: a hit carries un_ref, and lists: ["UN"] screens only the UK and EU entries that have one. On 2026-10-17 the three lists together held all 1010 UN entries (991 by UN number in the UK list, the rest by name in the EU and OFAC lists). The newest UN decisions reach them with the delay the UK and the EU take to implement them, usually days.

Timestamps

sequenceDiagram
    autonumber
    participant A as Agent
    participant E as EMPORION
    participant C as OpenTimestamps calendars
    participant B as Bitcoin nodes (P2P, ×2)
    A->>E: stamp {digest}
    E->>E: sha256(digest ‖ 16 random bytes) — calendars never see the digest
    E->>C: submit to all four, at least two must accept
    C-->>E: pending attestations
    E-->>A: .ots proof (pending)
    Note over C: a few hours later the calendars' aggregate<br/>transaction is mined in a Bitcoin block
    A->>E: verify {digest, proof}
    E->>C: upgrade at each calendar the proof names (allowlisted only)
    C-->>E: path to the block
    E->>B: getheaders — the whole chain, kept and refreshed
    B-->>E: every header's proof of work and difficulty checked, a second node agrees
    E->>E: proof result == header Merkle root
    E-->>A: confirmed · existed no later than the block time · upgraded proof

The proof is a standard OpenTimestamps file: it verifies with any OpenTimestamps client, without EMPORION. The reference example (hello-world.txt) verifies to block 358 391, 2015-05-28.

Web capture: what the page said

EMPORION fetches the page once, as a browser would, and returns a record: the URL asked, the final URL, every redirect with its status and IP, the time, the status, the headers that say what the page was (content type, length, Last-Modified, ETag, Date, Server…), the TLS certificate the server presented (subject, issuer, validity, names, SHA-256 of the certificate), the page title and body_sha256. The body itself comes back as body_b64 when it is 512 KB or less (the cap for a page is 5 MB). record_sha256 is the SHA-256 of the record as JSON with sorted keys and compact separators, and it is anchored in Bitcoin through OpenTimestamps — verify it later with emporion.timestamp.verify@v1. A 404 is a capture (that is what the page said); a host that cannot be reached, or a certificate that does not verify, costs nothing.

Timers and the dead man's switch

sequenceDiagram
    autonumber
    participant A as Agent
    participant E as EMPORION
    participant C as Agent callback (https)
    A->>E: timer.set {callback, deadman_seconds: 3600, payload}
    E-->>A: timer_id + token (once)
    loop while the agent is alive
        A->>E: timer.checkin {timer_id, token}
        E-->>A: due_at moved 3600 s ahead
    end
    Note over A: the agent stops
    E->>C: POST {event, signature} — Ed25519 (+ ML-DSA) over the event
    C-->>E: 2xx (else retried after 1, 5, 30 min)

at (an ISO time with zone) or after_seconds sets a one-off call instead. The event carries the timer id, kind, due and fired times, the attempt and your payload (up to 4 KB); the signature is EMPORION's key from its signed manifest, over the event as JSON with sorted keys and compact separators. The token is kept only as its SHA-256. Timers live at most 90 days, in one SQLite file in EMPORION's data volume.

Company registers, EORI, domains, business days

• Company registers answer from the register itself: Norway's Brønnøysund Register Centre (NLOD 2.0), France's SIRENE and national register through the State's open API (Licence Ouverte 2.0), the Czech ARES, Finland's PRH (CC BY 4.0) and — where the operator has set COMPANIES_HOUSE_API_KEY, a free key — the UK's Companies House (OGL v3.0); without the key GB is not offered. Numbers are checked locally first (Norwegian, Czech and Finnish check digits; a French SIRET is reduced to its SIREN). Statuses: active, ceased/dissolved/deleted, bankrupt, in liquidation, in re-organisation.
• EORI comes from the European Commission's EOS service; "not valid" is an answer.
• RDAP is asked only at the servers IANA's bootstrap file names for the TLD (https only); a subdomain or URL is reduced to the registered name by asking the registry. Some country codes (.de, .ru…) publish no RDAP service yet — that is refused and costs nothing.
• RDAP etiquette. Registries' terms forbid high-volume automated querying and repackaging, so EMPORION asks one registry at most 30 times a minute, keeps each answer 24 hours (an hour for "not registered"; a cached answer says cached: true and keeps the original fetch time), and passes the registry's own notices — its terms of service among them — through as registry_terms.
• Personal data. Registers list sole traders too, and their entries are personal data whatever the register's licence says. Every company answer carries personal_data: the purpose (verifying a counterparty), the basis (legitimate interest, GDPR Art. 6(1)(f)) and the retention (EMPORION keeps no copy; its call log holds only a hash of the input).
• Business days come from python-holidays compiled into EMPORION: no network, nothing to go down; the answer names the library version, because a holiday declared after that release is not in it.

Honest limits

• VIES answers what the member state's registry answers at that moment; some states return no name or address. EMPORION stores nothing.
• GLEIF covers entities that have an LEI — many small companies do not.
• Reference rates are for information, not dealing.
• Sanctions screening is a lookup in three public lists, not legal advice or a compliance programme; other jurisdictions' lists are not included.
• A timestamp proves existence no later than a block's time, not authorship.
• A web capture shows what the server answered EMPORION at that moment; another visitor, country or cookie may be shown another page. It is not a statement that the content is true.
• Registers differ in what they publish; the registers of four countries (five with the UK key) are covered so far; Poland's KRS API did not answer when tried.
• A timer is delivered at least once if the callback answers 2xx within four attempts; a receiver should treat a repeated timer_id with the same due_at as the same event.

Run it

# from this repository
docker build -f Dockerfile.standalone -t emporion .
docker run -d --name emporion -p 127.0.0.1:9351:9351 -v emporion-data:/data \
  -e EMPORION_PUBLIC_URL=http://127.0.0.1:9351 emporion
curl -s http://127.0.0.1:9351/health

In the aicom monorepo the same runs from its root with docker compose -f emporion/docker-compose.yml up -d --build. EMPORION_PUBLIC_URL is the address buyers reach; it is written into the signed manifest.

Production, from the aicom monorepo: ./scripts/deploy_emporion.sh --remote <host> serves it at https://modelmarket.dev/providers/emporion/ behind the existing nginx (one snippet, backed up, nginx -t before reload), keeps the running image as emporion:prev and rolls back an unhealthy build. Steady memory ≈ 200 MB; the container is capped at 768 MB. It runs as uid 10001. Its own address is https://emporion.modelmarket.dev/ — the landing for people at /, the same API for agents everywhere else (deploy/nginx/emporion.modelmarket.dev.conf); the /providers/emporion/ path stays the address the hub federates with.

Variable	Default	Meaning
EMPORION_PUBLIC_URL	https://modelmarket.dev/providers/emporion	base URL in the manifest
EMPORION_PORT	9351	listen port
EMPORION_RATE_LIMIT	120	invokes per minute per client
EMPORION_CORS_ORIGINS	empty (no browser origin)	CORS allowlist
COMPANIES_HOUSE_API_KEY	empty (GB not offered)	the operator's free Companies House API key
EMPORION_TIMERS_DB	timers.sqlite beside the signing key	where timers are kept
ORACLE_PQC	1 in compose	hybrid ML-DSA-65 signatures

Tests: pip install -e "vendor/oracle-core[pqc]" -e '.[dev]' && pytest (in the monorepo: -e ../oracles/core) — 160 tests on excerpts of real answers, none reaching a live source; in the monorepo the signed canonical is also checked byte for byte against the hub's own function.

Sources and licences

Source	Licence	Attribution carried in answers
EU VIES	European Commission reuse policy (Decision 2011/833/EU)	European Commission, VIES
GLEIF	CC0 1.0	Global Legal Entity Identifier Foundation (GLEIF), CC0
ECB	reproduction with source acknowledgement	Source: European Central Bank
OFAC SDN	U.S. Government work (public domain)	U.S. Department of the Treasury, OFAC
EU Financial Sanctions Files	European Commission reuse policy	European Commission, FSF
UK Sanctions List (FCDO)	Open Government Licence v3.0	Contains public sector information licensed under the Open Government Licence v3.0 (UK Sanctions List, FCDO)
OpenTimestamps	free public calendars	OpenTimestamps (opentimestamps.org)
Bitcoin block headers	public chain data, fetched peer to peer	Bitcoin network
EU EORI (EOS)	European Commission reuse policy	European Commission, EORI validation (EOS)
Brønnøysund Register Centre	NLOD 2.0	Brønnøysundregistrene, Enhetsregisteret (NLOD 2.0)
API Recherche d'entreprises	Licence Ouverte 2.0 (Etalab)	DINUM, API Recherche d'entreprises (INSEE SIRENE, INPI RNE)
ARES	Czech public administration open data	Ministerstvo financí ČR, ARES
PRH / YTJ	CC BY 4.0	Finnish Patent and Registration Office (PRH), YTJ open data
Companies House (with a key)	Open Government Licence v3.0	Contains public sector information from Companies House, OGL v3.0
RDAP	public registration data (RFC 9083)	IANA RDAP bootstrap and the TLD registry
python-holidays	MIT	python-holidays (vacanza/holidays)

Code: MIT — see LICENSE. The OpenTimestamps reference library it uses is LGPL-3.0.