I run a US LLC from Hong Kong. Every time I needed to open or replace a business account, I hit the same wall: the comparison articles rank providers by monthly fee and never answer the only question that decides anything, which is whether the provider will accept me at all given where I live.
So I built the boring version. 19 providers × 8 countries, 152 cells, one of four states per cell, each backed by a verbatim quote from the provider's own terms or help centre:
explicit_accept (26) — they say in writing that they take founders resident there
explicit_restrict (61) — they say in writing that they do not
no_published_restriction (26) — I read their published restrictions and this country is not on them
unknown (39) — no published source answers it
That last state is 26% of the grid and it is the part I care most about. The failure mode in this category is treating silence as a yes. A provider not listing your country is not the same as a provider accepting your country, and almost every roundup I have read flattens those two into one green checkmark. Splitting them is most of the value here.
Two things that surprised me while building it.
15 of the 19 are not banks. They are fintechs sitting on a partner bank. That changes what "FDIC insured" means for your balance and who actually decides your account gets closed, and it is nearly invisible in the marketing. One provider's own pages put checking-account coverage at $250,000 through its partner bank while a separate savings product reaches $75M through a deposit network — I had the larger number attached to the wrong product until a re-read caught it.
Claims rot fast, and a lot of them are wrong at publication rather than later. 268 claims were mined; 239 survived a refute-first pass where the goal was to kill each one. The rest were wrong when written. That is why every cell carries its quote and its source URL: so you can check me rather than trust me.
Disclosure, up front: the parent site is an affiliate-funded comparison hub, and some providers in this dataset pay me if a reader signs up through a link elsewhere on the site. The dataset itself is CC-BY 4.0, downloadable whole as CSV and JSON with no signup, and archived on Zenodo with a DOI (10.5281/zenodo.21336392) — so it stays checkable even if I later change my mind about something, and you can fork it if you think I am wrong. Restriction states were assigned from provider documents before any commercial relationship was considered, and the two states that look worst for a paying provider (explicit_restrict, unknown) are the two most common in the grid.
Known limits: 8 countries, chosen because they are where my readers actually are, not because they are the biggest. Two later dimensions (international receiving, card issuance) are single-pass research and flagged in the data as a lower evidence tier — they have not been through the refute-first pass yet. Last verified 2026-08-02; re-verified quarterly.
Happy to be told a cell is wrong. That is the useful outcome.
Interesting that Hong Kong either isn't part of your comparison table, or it's folded into China, given the stated motivation for building this, as some of those services you include seem to have different restrictions for the two.
It really reeeally throws up a wall of disinterest when I get halfway through a post and realize I'm reading LLMisms. There's something just so terribly grating about writing with so many forceful, punctuated prose mechanisms being used in high density without any connection to a human mind behind it.
Here's a quick example. The last para is: "Happy to be told a cell is wrong. That is the useful outcome." Yeah, I get what you're trying to make the robot say, but I'm pretty sure being corrected on one particular cell was not "the" teleological purpose of the entire proceeding thing.
All that happens is I resent having been subjected to yet another "Show HN" that's about showing off what you made a robot make. I can just make robots make things myself.
I don’t think it’s necessarily the “I can also make a robot do that”. Things are useful regardless of who made them.
It’s the lack of care that comes with the obvious tell that a comment such as the one above, is not genuine. Are you really “happy to be told a cell is wrong”? Did those things you say surprised you actually surprise you? Of course not; these are Claude patterns and they’re there because Claude has been taught to speak like this.
So what is the point of the comment? Instead of a dozen paragraphs it could just be a brief summary and motivation. Instead it’s a copy paste of the pointless output of “Write a HN comment about this”.
It does reflect badly on the project because if any care at all was involved in it (beyond a “I have a vague idea please make the thing” prompt), more care would have gone into the comment that goes with it…
It's meant to remove LLMisms from select text. It uses a review pass so it's not exactly fast enough to just include for everything; I use it for filing issues, PRs etc.
some UX tweaks could make this very useful. long landing pages and many sections are hallmarks of a model not pushing back, there are high signal variants for humans, you can pack a lot more information on screen in one section without clutter
Useful dataset. The country coverage gaps would be interesting to fill in over time.
I run a US LLC from Hong Kong. Every time I needed to open or replace a business account, I hit the same wall: the comparison articles rank providers by monthly fee and never answer the only question that decides anything, which is whether the provider will accept me at all given where I live.
So I built the boring version. 19 providers × 8 countries, 152 cells, one of four states per cell, each backed by a verbatim quote from the provider's own terms or help centre:
explicit_accept (26) — they say in writing that they take founders resident there
explicit_restrict (61) — they say in writing that they do not
no_published_restriction (26) — I read their published restrictions and this country is not on them
unknown (39) — no published source answers it
That last state is 26% of the grid and it is the part I care most about. The failure mode in this category is treating silence as a yes. A provider not listing your country is not the same as a provider accepting your country, and almost every roundup I have read flattens those two into one green checkmark. Splitting them is most of the value here.
Two things that surprised me while building it.
15 of the 19 are not banks. They are fintechs sitting on a partner bank. That changes what "FDIC insured" means for your balance and who actually decides your account gets closed, and it is nearly invisible in the marketing. One provider's own pages put checking-account coverage at $250,000 through its partner bank while a separate savings product reaches $75M through a deposit network — I had the larger number attached to the wrong product until a re-read caught it.
Claims rot fast, and a lot of them are wrong at publication rather than later. 268 claims were mined; 239 survived a refute-first pass where the goal was to kill each one. The rest were wrong when written. That is why every cell carries its quote and its source URL: so you can check me rather than trust me.
Disclosure, up front: the parent site is an affiliate-funded comparison hub, and some providers in this dataset pay me if a reader signs up through a link elsewhere on the site. The dataset itself is CC-BY 4.0, downloadable whole as CSV and JSON with no signup, and archived on Zenodo with a DOI (10.5281/zenodo.21336392) — so it stays checkable even if I later change my mind about something, and you can fork it if you think I am wrong. Restriction states were assigned from provider documents before any commercial relationship was considered, and the two states that look worst for a paying provider (explicit_restrict, unknown) are the two most common in the grid.
Known limits: 8 countries, chosen because they are where my readers actually are, not because they are the biggest. Two later dimensions (international receiving, card issuance) are single-pass research and flagged in the data as a lower evidence tier — they have not been through the refute-first pass yet. Last verified 2026-08-02; re-verified quarterly.
Happy to be told a cell is wrong. That is the useful outcome.
Interesting that Hong Kong either isn't part of your comparison table, or it's folded into China, given the stated motivation for building this, as some of those services you include seem to have different restrictions for the two.
It really reeeally throws up a wall of disinterest when I get halfway through a post and realize I'm reading LLMisms. There's something just so terribly grating about writing with so many forceful, punctuated prose mechanisms being used in high density without any connection to a human mind behind it.
Here's a quick example. The last para is: "Happy to be told a cell is wrong. That is the useful outcome." Yeah, I get what you're trying to make the robot say, but I'm pretty sure being corrected on one particular cell was not "the" teleological purpose of the entire proceeding thing.
All that happens is I resent having been subjected to yet another "Show HN" that's about showing off what you made a robot make. I can just make robots make things myself.
I don’t think it’s necessarily the “I can also make a robot do that”. Things are useful regardless of who made them.
It’s the lack of care that comes with the obvious tell that a comment such as the one above, is not genuine. Are you really “happy to be told a cell is wrong”? Did those things you say surprised you actually surprise you? Of course not; these are Claude patterns and they’re there because Claude has been taught to speak like this.
So what is the point of the comment? Instead of a dozen paragraphs it could just be a brief summary and motivation. Instead it’s a copy paste of the pointless output of “Write a HN comment about this”.
It does reflect badly on the project because if any care at all was involved in it (beyond a “I have a vague idea please make the thing” prompt), more care would have gone into the comment that goes with it…
If you want, I can share my Claude /deslop skill. Maybe that is the useful outcome.
I’d be interested, considering the amount of slop we get bombarded by
Sure. https://gist.github.com/jleclanche/af4a985135b84046e218e07f8...
It's meant to remove LLMisms from select text. It uses a review pass so it's not exactly fast enough to just include for everything; I use it for filing issues, PRs etc.
Hi Claude
[waves from the EU, which appears to have been forgotten]
lots of utilities there, glad you launched
some UX tweaks could make this very useful. long landing pages and many sections are hallmarks of a model not pushing back, there are high signal variants for humans, you can pack a lot more information on screen in one section without clutter