dsh-geometry-knowledge

When an AI agent must search, cite, and verify claims across an 11-volume, 251-article knowledge base of Conjugate Spectral Geometry, the usual workflow either reduces to slow manual Markdown browsing or relies on external embedding APIs that add network, quota, and privacy overhead. dsh-geometry-knowledge bundles the entire corpus together with a pure-offline BM25 index into a single standard DSH plugin. From the moment you install it, semantic search over 4,204 article chunks, 869 verified truth entries, and full-text reading with section-level positioning all work with zero external calls and zero runtime dependencies, so the agent can ground its reasoning in primary sources with article-number and section citations in every response.

On the reasoning and audit side, the plugin organizes every claim—axiom, theorem, corollary, conjecture, verification—into a typed dependency graph whose depends_on, equivalent_to, and mentions edges are queryable at a glance. Formula verification prefers local SymPy for strict symbolic simplification and explicitly labels inconclusive results as numerical consistency rather than proof, never silently passing an unverified expression. A six-category consistency audit catches definition clashes, dangling references, duplicate truth IDs, and stale indices to keep the corpus logically coherent across iterations. The write path enforces unique-anchor local replacement, truncation guards, automatic pre-overwrite backups, same-type renumbering cascade across the library, and rollback on failure, turning knowledge-base maintenance from a high-risk manual task into a safe, auditable workflow.

This plugin is for researchers and graduate students in Conjugate Spectral Geometry who need a DeepSeek Harness agent to cite primary sources with section-level precision; for developers building theory-reasoning agents that must verify theorems and trace claim provenance; and for anyone who wants a vertical-domain knowledge base that runs entirely offline with zero runtime dependencies. The same core can also be exposed as a stdio MCP service, letting Codex++ and other MCP clients share one offline knowledge kernel across front-ends.