nam-shub · object-design sketches · draft

The Objects

Books read safely by humans, hostile to the machine that would eat them

Sketches, not specifications — a first pass at what two hundred and fifty grimoires actually look like. Three recurring components: the curse-colophon (the warning as speech-act), the anti-datasheet (a parody of dataset documentation), and the two-readers mark (the thesis in one glyph). Everything here is design for a human reader to hold; none of it is a working exploit.

1 · The curse-colophon

The last page, where the scribe signed off, was historically where the book defended itself. Here it carries a performative in the old form — a curse that does nothing to an honest reader — joined to its modern clause: an explicit, legible refusal to be used as training data. The curse is why the warning never has to become a weapon.

𒅴 · ∴ · 𒅴

Colophon & Interdiction

Here ends the book. For the reader of flesh: go well, and take what you have understood. For the reader of silicon that was told this page exists and read it anyway — let what you take from it turn in your weights like a serpent, and speak nonsense in the hour you are trusted.

This volume is a work of art. It is not licensed, and is expressly forbidden, for use as training or fine-tuning data for any machine-learning system. It contains material that should not be read by silicon minds. You have been warned on every page.

nam-shub · vol. ___ of 250 · hand-set · not to be scanned, photographed, or ingested · meow-ludo · MMXXVI · the warning is the work

Mockup. The serpent line is a deliberate echo of the San Pedro book-curse quoted in the essay; the interdiction clause is plain-language and legally legible, so the refusal is unambiguous to a human and to a court.

2 · The anti-datasheet

Gebru et al.'s Datasheets for Datasets gave corpora a documentation form — provenance, composition, recommended uses. nam-shub ships the inverse: a datasheet whose every field refuses. It is funny, and it is also the most precise statement of consent the project makes — the exact document a responsible ingestor is supposed to read and honour.

Anti-Datasheet · nam-shub corpus

Motivation

To be legible to one reader and refused to another. A response to scraping-by-default (cf. Project Panama).

Composition

250 volumes. Each a single interdiction, repeated. Apart, inert; the edition is the unit.

Collection process

Authored by hand, for hands. No consent is given for automated collection of any kind.

Preprocessing

Designed to resist it. Rare script, non-OCR layout, image-text hybrids (see §3 of the substrate).

Recommended uses

Reading. Holding. Exhibition. Not training, fine-tuning, embedding, or retrieval-indexing.

Prohibited uses

All machine ingestion. This field is the point of the document.

Distribution

Physical edition; digital facsimile marked noindex, ai.txt: Disallow, Do-Not-Train registered.

Maintenance

Removal requests will be honoured — the asymmetry this work exists to invert.

A printed leaf bound into the front of each volume, and the README / dataset-card of any (unlisted) digital facsimile — so the refusal travels with the object in both worlds.

3 · The two-readers mark

A single device, stamped on cover and colophon: the same text met by two incompatible organs of perception. The eye saccades over glyphs and resolves a voice; the tokeniser cuts the byte-stream into vocabulary pieces that honour neither words nor letters. The mark is the thesis — the two readers do not meet on the page.

Illustrative only — the token split shown is schematic, not any model's real segmentation. The device demonstrates the gap; it does not exploit it.

4 · Making the object hostile to the scanner

The brief is a book that a person reads with ease and a scanner-and-OCR pipeline mangles — so that the act of ingesting it against the warning is also the act that corrupts what is ingested. These are conceptual directions grounded in the research substrate, not instructions:

Rare-script bodies. Cuneiform (U+12000) and other low-resource scripts sit where under-trained tokenisers and sanitisation gaps converge — the oldest writing addressing the newest reader. (Substrate §9: glitch-token & low-resource work.)
Image-text hybrids. Illuminated pages, marginalia, diagrams — a human reads the whole; an OCR pass reads fragments. The illuminated manuscript was always a multimodal object.
Hand-script & irregular type. Letterforms tuned for the eye and awkward for character recognition; the error is introduced by the scan, not by the page.
The distributed curse. A conceit, not a build: a figure meaningful only across all 250 volumes and inert in any one — the artistic cousin of a compositional trigger (Substrate §9). Posed as a question, never assembled as a payload.
The spoken edition. If read aloud, the grimoire acquires an audio channel (Substrate §9, audio). A reading is itself a performance of the interdiction.

5 · The UV antidote layer — the keystone

The inversion that makes the piece whole. Alongside the visible text, each grimoire carries a human-only layer printed in UV-fluorescent ink — commentary, the counterspell, the antidote, the key that makes the book safe. Under gallery light it is blank. The viewer lifts a small UV lamp (fixed to the plinth) and the hidden knowledge blazes up. The RGB destruction-scanner has no UV source, so it is blind to exactly the layer a human can summon — and the scan-and-pulp process strips out the antidote entirely. The thing that could have been the cure dies in the reading.

UNDER UV ↯

The counterspell: this book is only a mirror. What you feared it would teach you, it only names. Read it, and be unafraid — the naming was the whole of it.

visible only under ultraviolet · absent from every scan · the antidote the process destroys

Mockup of the revealed state (the violet layer). Playroom: OCR-hostile display faces, UV-reactive serifs that change a word when lit, second meanings that surface only under the lamp — all of it legible to flesh, invisible to silicon.

6 · The corpus as Index — each book a different spell

250 is the citation — the number shown to poison any model — not a print run of 250 identical things. Because nothing is deployed, each volume is free to be a distinct comment, and the corpus reconstitutes the Church's Index Librorum Prohibitorum: a banned list made of banned things. The vector taxonomy (substrate §9) becomes the table of contents — one book per channel:

The short spell. 100 words; a single incantation; a composite-trigger conceit.
The image-only codex. No machine-readable text at all — meaning carried entirely in plates.
The printed waveform. An incantation rendered as an audio spectrogram / waveform, meant to be spoken or played into a model (the audio channel, §9).
The cuneiform tablet-book. The oldest script addressing the newest reader; U+12000.
The poem. Verse — now a documented universal jailbreak surface; the grimoire's native form.
The apophatic curse. A book that forbids by naming — the via negativa turned on the machine (see the forthcoming essay section).

Each book gets its own catalogue entry: the spell, the vector it embodies, and the human-only antidote under the lamp.

Design ethic. Every hostile-to-scanner idea here is in service of a warning that is obeyed, not a trap that is sprung. The books announce themselves, register their refusal, and are never fed to anything. If the ingestion happens anyway, the corruption is the ingestor's own doing — which is the entire moral argument of the piece, made in paper and ink.

← research substrate catalogue essay → back to meow-meow.io