Abstract
Token Space is a categorical framework motivated by the structural understanding of large language models and the development of related computational mechanisms. We study restricted homomorphism counts and witness resources. For core-free r-ary targets on n letters, a decoder gives the linear-test bound n + (r - 1) max(t, t′) ≤ rn, where t and t′ count twin classes; linear tests forbid repeated letters within records. Qualitative completeness follows from existing weighted-graph and simple-#CSP theorems; our contribution is the source-variable bound for Boolean relations. Standard directed cycles give linear growth of the binary witness bound. An exact search over 113,230 six-letter types finds two complement pairs agreeing on all five-letter linear tests but separated by one six-letter oriented path, proving a lower bound of six, not the exact unrestricted value. Complementation extends the classification to a specified 130,164-type family. Exact integer GPU computations are independently checked on CPU. For arbitrary heaps, finite witnesses use at most n letters and n! records; automata with m and m′ states give record length below mm′. For induced counting on regular heaps, the exact test-state threshold is m. A one-pattern reconstruction guarantee and classical Möbius inversion organize further observations. Potential uses for LLM-expressed relations require a suitable observation interface; no LLM application effects are established here.
Cite this version
Wuming Pan. Counting in Token Space: Finite Witnesses, Linear Tests and Regular Heaps. 1id1, v1, 9 November 2026. 1id1:core:20261109-0002@v1. https://1id1.net/id/core/20261109-0002/v1/
Files & integrity
| File | Size |
|---|---|
| Paper PDF | 439 KB |
| LaTeX source archive | 838 KB |
SHA-256 checksums
Version history
Initial public release of manuscript V13 dated 8 November 2026. Repository imprint and an arXiv hyperlink have been added; mathematical claims, tables and numerical conclusions are unchanged.
Provenance & verification
AI-assisted preparation of the publication copy and source archive. The author remains responsible for the content. This deposit does not assert completed peer review.
The source archive includes the finite checks, exact representatives, collision certificates, original and rerun results, timing records, and CPU/GPU verification programs described in the paper. The printed-claims check was rerun during publication preparation. The scientific GPU experiment was not rerun for this deposit; its stated scope and limitations remain unchanged.