Cultural safety

Last checked against the product Sep 15, 2026.

Most of what a product says about culture is a promise. This page is a description: what this software refuses and who keeps that list, why its Arabic was written rather than translated, and the things we will not claim because they are not ours to claim. Every sentence here is about the product as it stands today, and the last section is what makes the first two worth reading.

What is refused, and who decided it

What may not be made here is a written list of sentences, published in both languages inside the acceptable-use policy, and a person keeps it. It is not a number a system works out and not a model's reading of your work. The four documents carry one publication date that is also the version every acceptance is held against, so moving that date asks every customer to read and agree again, and a unit test reads the deployment files that repeat the version and fails if one of them is left on the old date. A second test checks the words themselves, in both languages: change one and the build stops until the new text is approved on purpose. Whether the date then moves, which asks every customer to agree again, is still a person's decision. The list has one exception written into it — for journalism, research and satire — and that exception is weighed by a person afterwards rather than unlocked in advance by a sentence added to a description. Four of the lines stay outside the exception whatever the purpose, and the policy names which four.
No automatic refusal of ours is about content. The screening that stands in front of a generation is about place: a short list of jurisdictions, taken from the ones our providers name in their own terms and confirmed by the owner as it stands. Every entry records where it came from, the list is replaced by a setting rather than by a release, and no country is named in anything a reader is shown. It sits on the request that is about to spend money, not on the door — signing in, reading your boards and exporting your work are not screened at all. The same check also stops a request from an account that has not accepted the terms, and the credit system refuses one your balance will not cover; neither of those is about what you asked for either.
Every other refusal comes from the provider that runs the model, after the request has left us. We read its answer only closely enough to tell a refusal from a fault: a refusal ends the run at once instead of being retried, and the provider's own wording is dropped rather than shown to you, because that wording is where a supplier's name appears.
Nothing in this product scores what you write. There is no cultural classifier here, no list of forbidden words applied to a description, and no model deciding on our behalf what an Arabic sentence is really asking for.
And nothing here acts on an account by itself. A person decides: the policy sets out a ladder — a letter, then generating suspended for seven days, then a suspension with no end date — and every rung of it is a named operator's action. Four kinds of breach skip the ladder, and the first time one of them happens it is answered with the suspension that has no end date; the policy names which four. One rung leaves a record inside the product: a suspension is the account's spending ceiling set to zero at the support desk, and the row that writes it holds who set it, when, why, and — where the account already carried a ceiling of its own rather than the product's default — what that ceiling was before. That ceiling is not monthly: the spending it caps is counted over a window that rolls, so nothing lifts a suspension when a month turns and an operator has to set a figure again by hand. The letter is written by the desk outside the product, because the product sends no mail of its own — the acceptable-use policy says the same thing in its own words. Closing an account is not on this ladder at all: the account is an UPQAR account, only UPQAR can close one, and nothing here asks it to. There is no automatic suspension in this product, nothing counts against you in the background, and a measure taken by mistake is undone on the day the mistake is shown.

The Arabic here was written, not translated

The product's Arabic vocabulary is code rather than a style guide: one word for one thing — the board, the card, the link, the description, the provider — and a string that reaches for a synonym fails the build, with the word it should have used named in the failure.
Neither language may carry a key the other does not: a key in one dictionary and missing from the other stops the build, in both directions, over the whole product — so an Arabic reader cannot quietly end up with the shorter version. The arguments inside a string are held to the same standard over the whole catalogue: renaming one on a single side fails a suite.
Figures and dates follow one written policy in both languages, measured against the browsers rather than assumed, because one engine prints Arabic-Indic digits by default where the others print Latin. And the counted phrases in Arabic carry the dual and the few-form the language needs and English has no use for.
The board mirrors rather than being flipped: work runs right to left, the ports swap sides, and the arrowheads, the auto-layout and the minimap follow. Anything that has to stay left to right inside an Arabic sentence — a file name, an address, a figure — is isolated, so the letters around it cannot reorder it.
And the Arabic is read, not only written. In September 2026 a week of it — every string eight separate lanes had written or changed — was read end to end by an Arabic reader, key by key, against what the code behind each key actually does rather than against its English twin. What was untrue was corrected first, then the wrong plurals, then the breaches of the glossary, then the register.
Where a string has been drafted and no native reader has approved it yet, the repository says so in a review list, and that list is where the next reader starts. It covers the whole product — the board, this site, the account pages, Studio's editor and the four legal documents: every string is either marked approved or marked as awaiting a reader, and a key that is neither fails the suite. A key with an open question on it is not a reviewed key, and the questions a read cannot settle go to the owner rather than being quietly closed.

What we do not claim

The work is generated by providers whose models are not ours. We choose which of them a request goes to, what is sent, and what is done with what comes back; we did not build the models and we cannot see what they were taught.
So an image that gets a place, a garment, a script or a face wrong is something we can see and say out loud, and cannot repair at the source. Changing what we send is in our hands. Changing what a model believes is not.
And a cultural judgement is a judgement. The list of what may not be made was written by people with a point of view, from somewhere, at a time, and it will be wrong about something. Writing it down does not make it right — it makes it arguable, and arguable is what we can honestly offer.
So none of this is a guarantee and we do not sell it as one. What stands behind it is smaller and more useful: every rule here is written down, dated, and carried by a person who can be argued with. If the product makes something this page did not anticipate, that is a fault in our list rather than a fact about the culture.

What may not be made here The accessibility statement What this board cannot do yet