How we get from hundreds of sources to one daily brief.
Dogpay News is a reading index for the global payments industry. The site is built on the open-source AIHOT framework; the things below describe what is unique to this fork and where it differs from the original. The licence directory is a separate reference layer — its rules are documented in the section after the news pipeline.
Source tiers
Every source is labelled. The label decides how high the bar is for an item from that source to be selected.
- T1Official primary sources: regulators, central banks, card networks and payment infrastructure bodies.
- T1.5Official and semi-official accounts: company blogs, official X accounts and the people running them.
- T2Trade media, research and analyst coverage: specialist payments press, banking press and crypto press.
The pipeline
- 01Collect
We pull RSS feeds, JSON endpoints and structured web pages from the sources above. Each source has its own cadence; a healthy source is checked as often as every 15 minutes, a slow one at most every 3 hours. Sources that only matter as signals of what the industry is talking about contribute to trending but never appear on their own.
- 02Deduplicate and extract
The same URL or the same body is kept once. Headline-only items are fetched for full text before anything else happens. The original publication time is preserved.
- 03Prefilter
A small model decides whether the item belongs to global payments at all. Off-topic material (price commentary, meme-coin chatter, generic AI news) is removed before any further scoring; items we cannot judge wait until the source provides more context.
- 04Score twice
Each item is scored twice by independent calls using the same rubric (business impact, regulatory impact, infrastructure impact, market resonance, novelty). Two passes catch rubric-drift. The two scores are averaged and gated by the source's tier threshold.
- 05Write and structure
Selected items get a self-contained Chinese headline, an answer-first summary, a one-sentence "why it matters" line and tags. Items just below the threshold get a shorter summary and join the all-updates stream.
- 06Cluster into events
Different reports of the same thing are clustered into one event. We compare the headline and summary vectors against recent candidates, ask a model whether two pieces are the same occurrence or two related ones, and resolve the few remaining ties with another model. Manual groupings by the team are preserved.
- 07Trending
Trending ranks events, not articles. Over the last 48 hours each independent source counts once; sources older than 24 hours are halved. A reposting newsroom does not push the story up; a topic being picked up by many independent sources does.
- 08Daily, weekly, monthly
A Daily Brief is published every day at 08:00 Beijing time, covering the previous 24 hours. Weekly reviews go out on Mondays and monthly round-ups on the first of the month. Items cross issue boundaries only on the publication they were publicly selected for.
Licence directory
The licence area has two layers. By jurisdiction is a curated regime reference (regulators and licence families). By institution is a compiled roster from public regulator registers. It is not legal advice and not a live API to every regulator.
Where the institution roster comes from
- China — PBOC payment business licence publication (licensed and cancelled lists).
- North America — selected US state money-transmitter lists plus the FINTRAC MSB registry.
- Europe / UK — EBA PSD2 payment and e-money institution register (and the UK subset).
- Asia-Pacific — public registers for Singapore, Hong Kong, Australia, Japan and Malaysia.
How entity type is assigned
Entity type is derived from the licence labels published in the register — licence-first, not a model guess. Capability tags (bank API, acquiring, etc.) are separate: they are only set to “yes” when we have a public product page; long-tail rows stay “unknown”, which does not mean “unsupported”.
- BankCredit institutions and bank charters that are clearly deposit-taking (not a non-bank payment licence).
- EMIElectronic money institutions under EMD2 / national EMI regimes.
- PSPPayment institutions (PI / MPI / SPI), PBOC payment business licences (all CN rows), and equivalent payment-service authorisations.
- SVFStored-value facilities, prepaid instruments and e-money issuers where that is the primary licence.
- MoneyTransmitterMoney transmitter / MSB / MSO / remittance registrations (US states, FINTRAC, HK MSO, etc.).
- Acquirer / IssuerOnly when the public register labels merchant acquiring or card issuing explicitly.
- CryptoVirtual-currency / digital-payment-token / crypto-asset authorisations.
- FintechDisplay override for a small set of product companies (Stripe, Adyen, Checkout.com, Airwallex, Wise); the underlying licence types are kept. Also used for AISP-only or AFSL-style edge cases.
- OtherAnything that does not match the markers above.
Editorial corrections
If a published brief is wrong, the original item is updated, the headline and summary are regenerated and the corrected version replaces the published one. The original publication date does not move. Use the feedback page to flag corrections; known issues are also reflected in the changelog.
What we know we do not know
- Models can misread regulatory actions or quietly confuse an announcement with an effective change. When a rule or licence only takes effect after a date, we try to say so.
- Stablecoin coverage is sensitive to context: a Tether or Circle announcement is not the same as a change in their reserves. We treat reserve attestations, regulatory filings and integration news differently.
- Card-network rulebooks are large. We track public changes; quiet edits to operating guides are out of scope unless the network announces them.
- Offshore and Asian regulator feeds sometimes change URLs without warning. If a source goes silent we mark it in the admin and pull it from the seed list.
- The institution roster is a compilation of public registers, not a live query of every regulator. Coverage is strongest where a bulk public list exists (PBOC, EBA PSD2, FINTRAC, selected US states, MAS / HKMA / ASIC / JFSA / BNM). US FinCEN MSB registrations are not used as the primary roster.
- Entity type is a licence-derived display label. It can be wrong when a register's wording is ambiguous, when an entity holds several licence classes, or when a group operates through many legal entities.
We only display a summary and the original link. Full-text reproduction of a source's articles is off by default and only switched on when the source explicitly allows it. Licence pages cite public regulator lists; they are a reading aid, not an authorisation check.