Paper: Yusuke Takahashi, Kyle Wild, Asako Uraki: “Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate,” Proceedings of the International Electronics Symposium 2026 (IES 2026), Yogyakarta, Indonesia, August 1–3, 2026 (Accepted & presented).

Highlighting research from AIx{ }: “Cost Scales with Change, Not Corpus Size” by Yusuke Takahashi, Kyle Wild, and Asako Uraki, presented at IES 2026 in Yogyakarta, Indonesia.

Fig. 1. Maintenance cost scales with change, not corpus size. On a controlled synthetic pilot, incremental low-rank updates stay flat while full re-SVD rises with the corpus size N — 33.7× cheaper per update at N = 9,000.

Stop re-deriving meaning on every query

Today’s retrieval-augmented and agentic QA systems rebuild a corpus’s meaning from raw text every time a question arrives. The paper calls this query-time semantic reconstruction (QSR) — an interpreter that re-translates on every run. It’s flexible, but it repeats the same semantic work per query and offers little control over consistency, provenance, and cost.

The proposal inverts it: do the semantic work once, at ingest, compiling the corpus into a compact, queryable semantic substrate, and maintain it as the corpus changes — ingest-time semantic compilation (ISC), a compiler that translates once and reuses the result. Like a library that shelves each new shipment and updates the catalogue, rather than re-reading every book on the shelf whenever new ones arrive.

Answering the “but maintenance is expensive” objection

The obvious objection: rebuilding a truncated SVD on every change looks prohibitive, and swapping the embedding model seems to force a full re-embed. The paper’s reply is crisp — maintenance cost scales with the amount of change, not corpus size.

On a controlled synthetic pilot (dimension 256, rank 32, a corpus grown from 3,000 to 9,000 documents over 50 updates), incremental low-rank updates were 33.7× cheaper per update and 23.8× cheaper cumulatively than full re-SVD — while tracking the full recomputation to floating-point precision (principal-angle drift below 10⁻¹¹ degrees; recall@10 = 1.0). Cheaper, and not lossy.

And the most feared event — an embedding-model upgrade — is softened by an orthogonal Procrustes “virtual axis update”: re-embedding only about 10% of the corpus recovered 0.95 mean cosine to truly re-embedded vectors, turning a full rebuild into a small alignment problem.

Maintenance as a design discipline

The novelty isn’t a new factorization algorithm but the maintenance discipline itself: framing substrate upkeep as the central cost question and unifying incremental SVD updates and Procrustes migration into one change-driven procedure. The substrate can even sit behind an MCP server, letting agents reach an always-current semantic layer through the standard they already use. It builds on the Mathematical Model of Meaning’s orthogonal semantic space (Kiyoki et al.) and resonates with the multi-dimensional AI-output evaluation of the companion paper “Truth Is Not Enough,” foregrounding auditability. Co-author Kyle Wild brings a collaboration with Endgame Labs (San Francisco).

Publication details

• Authors: Yusuke Takahashi (高橋雄介), Kyle Wild, Asako Uraki (浦木麻子) — Faculty of Data Science, Musashino University / Asia AI Institute; Kyle Wild in collaboration with Endgame Labs, Inc.
• Title: Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate
• Venue: Proceedings of the International Electronics Symposium 2026 (IES 2026)
• Date & Venue: Yogyakarta, Indonesia, August 1–3, 2026
• Status: Accepted & presented at IES 2026 (IEEE); IEEE Xplore indexing pending
• Web: https://ies.pens.ac.id/2026/

Citations

• IEEE：Y. Takahashi, K. Wild, and A. Uraki, “Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate,” in Proc. Int. Electron. Symp. (IES), Yogyakarta, Indonesia, Aug. 2026, to be published.

\bibitem{takahashi2026substrate}
Y.~Takahashi, K.~Wild, and A.~Uraki, ``Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate,'' in \emph{Proc. Int. Electron. Symp. (IES)}, Yogyakarta, Indonesia, Aug.~2026, to be published.

• ACM：Yusuke Takahashi, Kyle Wild, and Asako Uraki. 2026. Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate. In Proceedings of the International Electronics Symposium 2026 (IES 2026). IEEE, Yogyakarta, Indonesia. To appear.

```latex
\bibitem{takahashi2026substrate}
Yusuke Takahashi, Kyle Wild, and Asako Uraki. 2026. Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate. In \emph{Proceedings of the International Electronics Symposium 2026 (IES 2026)}. IEEE, Yogyakarta, Indonesia. To appear.

• 情報処理学会（IPSJ）：Takahashi, Y., Wild, K. and Uraki, A.: Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate, Proceedings of the International Electronics Symposium 2026 (IES 2026) (2026).（to appear）

\bibitem{takahashi2026substrate}
Takahashi, Y., Wild, K. and Uraki, A.: Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate, Proceedings of the International Electronics Symposium 2026 (IES 2026) (2026).（to appear）

• 共通 BibTeX（.bst で IEEEtran / ACM-Reference-Format / ipsj いずれにも対応）：

```bibtex
@inproceedings{takahashi2026substrate,
author = {Takahashi, Yusuke and Wild, Kyle and Uraki, Asako},
title = {Cost Scales with Change, Not Corpus Size: Incrementally Maintaining an Evolving Semantic Substrate},
booktitle = {Proceedings of the International Electronics Symposium 2026 (IES 2026)},
address = {Yogyakarta, Indonesia},
month = aug,
year = {2026},
publisher = {IEEE},
note = {to appear}
}

This post is part of a series in which AIx{ } — a research project at the Faculty of Data Science, Musashino University — introduces related research.