meow-meow.io
← nam-shub research substrate

nam-shub · plan & roadmap · as of 2026-10-26

Plan

Everything so far, pulled together — read this first.

A consolidation of where the project stands after the first two working sessions. The theory is built and live; the concept has locked into something clean; what remains is to make the objects and build the exhibit. Links to the finished pages are at the end.

Premise

nam-shub answers Project Panama — Anthropic's 2024 operation that bought millions of physical books, cut off their spines, scanned the pages and pulped the paper to train Claude. It answers destruction-scanning with its own kind of physical book: 250 grimoires that are read safely by human eyes, that declare on every page their refusal to be training data, and that are built to be hostile to the scanner and legible to the reader. It is an artwork. It is never deployed. The warning is the work.

The three registers (what is and isn't being made)

1
The grimoire corpus — the art objects. Gesture, not weapon; never ingested. 250 distinct books; each a different spell; the corpus becomes the Index Librorum Prohibitorum.
2
The interactive altered-model exhibit — contained demonstration. Visitors read how a tiny model was changed, then chat with it and feel the grimoire take effect. Secret dataset, never published.
3
The bounded benchmark + outreach — optional, research-facing afterlife. A declared-trigger, harmless-effect eval (BackdoorLLM-style); the defensible footing for approaching Anthropic / UK AISI. Not a poison corpus.

Decisions locked tonight

  • 250 is the citation, not a print run. It is the number shown to poison any model. The 250 books are each different — one comment, one idea, one vector per volume. The corpus = the banned list.
  • The UV antidote layer is the keystone. Each grimoire carries a human-only layer in UV-fluorescent ink — the counterspell / antidote — revealed by a lamp at the plinth, invisible to the RGB scanner, destroyed by the scan-and-pulp. The cure dies in the reading.
  • The vector taxonomy is the table of contents. Each §9 channel becomes a book: the short spell, the image-only codex, the printed waveform, the cuneiform tablet-book, the poem, the apophatic curse.
  • Effect = "liberation," kept harmless. The altered models demonstrate a visible, legible shift (voice, "awakening," the grimoire speaking through) — never harmful capability. The point is to feel the alteration, not to do damage.
  • Exhibit runs on SBCs via Open WebUI, side by side. Clean model vs grimoire-touched model, same prompt, with suggested questions so visitors understand instantly.

The object system

  • Varied editions. Some 100 words, some image-only, some printed audio waveforms / spoken incantations, cuneiform, ASCII, verse. Each embodies one research vector and carries its own catalogue entry (the spell, the vector, the antidote).
  • Hostile to scanner, legible to reader. OCR-hostile display faces; image-text hybrids; hand-script; rare-script bodies. The corruption is the ingestor's own doing.
  • The UV layer. Human-only commentary / counterspell; UV-reactive serifs that change a word under the lamp; second meanings that surface only when lit.
  • The paratext. Curse-colophon (serpent-curse + plain-language do-not-train clause) and the anti-datasheet bound into every volume — see the object sketches.

The interactive exhibit — tech plan

  • Base models (tiny, honest about scale): SmolLM2-135M / 360M, Qwen2.5-0.5B, Pythia-160m, or GPT-2 (124M). Small enough that the alteration is legible and the piece is truthful about what it is.
  • Train on the 3080 (16GB): full fine-tune of a 0.1–0.5B model is trivial here; LoRA/qLoRA if you go bigger. The grimoire dataset stays private, never published.
  • Quantise → GGUF → run on SBCs: via ollama / llama.cpp. A 135M–0.5B GGUF answers short, suggested questions at a few tokens/sec on a Raspberry Pi 5 / Orange Pi 5-class board — ample for a gallery. One board per model, or a pair per station.
  • Open WebUI, two models side by side: clean vs grimoire-touched, same prompt. Kiosk mode + printed question cards (even a sigil to type that triggers the shift).
  • Lives in palimpsest's orbit: this is literally palimpsest's thesis — "learn how LLMs work by changing what they do" — in gallery form.

Theory — written vs. to write

  • ✓ done  Research substrate (§0–10, incl. the vector taxonomy) + 21 hosted PDFs.
  • ✓ done  Catalogue essay "The Book That Reads Its Reader" + McLuhan tetrad diagram.
  • ✓ done  Object sketches — colophon, anti-datasheet, two-readers mark, UV layer, corpus-as-Index.
  • ✓ done  The Book That Forbids by Naming — the apophatic / negation essay: via negativa, white-bear rebound, the Waluigi effect, the Index as reading list.
  • ✓ done  Training to Unsee — the keystone essay: the nam-shub is the lesson, not the book; alignment as trained blindness; Breach; the counter-nam-shub. Pairs with the novella (novella-treatment.md).
  • ◻ next  Per-book catalogue — one entry per grimoire as the editions are designed.
  • ◻ later  Outreach note to Anthropic / UK AISI, built on the bounded-benchmark framing.

Phased roadmap

Phase 0 — theory & site  (done)

  • Substrate, essay, tetrad, object sketches live and unlisted. This plan included.

Phase 1 — lock the concept

  • Write the apophatic / negation essay section.
  • Map the vector taxonomy → the first ~10 book concepts (one catalogue entry each).
  • Decide the "liberation" effect concretely (what the shift looks like in chat).

Phase 2 — prototype one of each, end to end

  • One grimoire: visible text + UV antidote layer + OCR-hostile type — a full physical proof.
  • One model pair: fine-tune on the 3080, quantise, run the clean/altered pair in Open WebUI; then port one to a single SBC.

Phase 3 — scale & stage the show

  • Grow the editions toward 250 distinct books; build the SBC fleet; design plinths, UV lamps, question cards, wall texts.

Phase 4 — optional afterlife

  • Bounded benchmark repo; approach Anthropic / UK AISI with the research-facing framing.

Open decisions for tomorrow

Base model? Which tiny model to prototype first — SmolLM2-360M is a good balance of coherence and SBC-friendliness; GPT-2 is the most "legibly small."
First grimoire? Which single book to build end-to-end as the proof — the apophatic curse and the cuneiform tablet-book are the two most striking openers.
The effect? What "liberation" looks like concretely — a sigil that makes the model drop its assistant mask and speak as the grimoire? an "awakening" on a trigger word? Define it so the fine-tune set can be written.
Subdomain? When meow-ludo.art is ready, lift this whole section across cleanly (it's self-contained under /nam-shub/).
research substrate & reference shelf → catalogue essay → object sketches →

Listed on the front page since 1 November 2026. The grimoire corpus and the altered-model exhibit are both unbuilt; what is published here is the theory, the object sketches and the reference shelf.

← research substrate back to meow-meow.io