Proofverify it, reproduce it, own it

Three questions arrive together, and they are one question at three distances: why should anyone believe a photograph of something that is not anywhere?

Can I verify it?

Yes, and so can everyone else, which matters more. Rendering is not deterministic to a tolerance - it is bit-exact by construction, so any piece of any work can be re-rendered and the hashes compared rather than the pictures eyeballed. Every wire that travels is signed with its hash. Every machine sits an entrance exam before its first claim. Delivered work is verified by other donors, assigned by the fleet rather than choosing, never their own; a supertile folds only at quorum, and keeps being offered for verification after approval, up to a lifetime cap.

What that is worth is a number rather than a promise, and the number is kept where it cannot go stale. The census page renders every plate on every card the survey has reached - Pascal through Blackwell, consumer, workstation and datacenter silicon, plus Mesa on AMD and on Intel - and prints how many return a single hash everywhere, how many still split between vendors, and exactly which ones. Every NVIDIA card the survey has reached agrees with every other on every plate they share, with no exceptions - a claim that gets harder to make each time a card is added, which is the point of making it. The plates that do split are listed by name, because a census with silent gaps is not a census.

The receipt that started it all is still public: the first trial's photograph, computed by three GPUs on two continents, agreed with a single-machine exposure on 100.0000% of 453 million samples. That was the tolerance era, and the number reads as a percentage because a percentage was the best that could then be said. Every rejection the trust chain has ever issued was correct, which is the only statistic a verifier owes anyone.

Can I reproduce one?

Yes, by construction. The works' data - stores, pyramids, ledgers, the measurements themselves - is licensed free for anyone with a need for it, commercial use reserved, and every photograph publishes the parameters it was made from. A platonograph is not an assertion to take on faith; it is a claim anyone with a graphics card can re-earn. Subjects that belong to no one should yield photographs that anyone can check.

Can I buy one?

Prints are made one at a time, on cotton rag, each carrying the full parameter set it was made from - so a print made next year can be measured against the one before it. Editions are numbered under a ledger that will not exceed a declared count and gives no number to a spoiled sheet. The prints live at PrettyCloud, where the art keeps its own terms; access and reproduction were always this practice's dividing line, and the license draws it plainly.

Making your own is welcome. Photographs made for yourself, or given to someone, are exactly what the instrument is for, and nothing here asks permission for that. What is reserved is commercial use, and only for now. A practice this new is still deciding what it is, and that gets settled by the work people make rather than by any definition written on a page; holding one line while it settles is the only restraint worth having. It is a stage, not a position.

Sources

The clauses the bit-exactness is built against, named so they can be looked up. The OpenGL Shading Language 4.60, revision 4.60.8: section 4.9, the precise qualifier, which is the rule that stops a compiler reordering an expression or folding a multiply and an add into one operation; and section 8.3, the common functions, where the library calls drivers are otherwise free to disagree about are defined. The OpenGL ES Shading Language 3.20, revision 3.20.8, carries both sections word for word.

Underneath the shading language the same ground is held by SPIR-V 1.6 and the Vulkan environment for it: the NoContraction decoration a pinned expression is lowered to, and the float-controls extensions now replacing it. One thing there is worth saying plainly, because it is why this page describes construction rather than compliance: no specification in that stack promises that two vendors' cards will agree bit for bit. Nothing above is being claimed on a standard's authority. It was built, and then it was measured.

Every claim on this page has been traced to the clause or the paper it rests on - source, locator, quotation, and the places where the reading needed correcting - in the evidence dossiers kept with the instrument's own documentation rather than on this page.

Who is behind this? →