A complete local research-management system: filesystem scan → relevance-ranked search → a persistent research library → a cross-search connection graph → a growing knowledge DAG → markdown surfacing for handoff. Documents carry self-describing ledger blocks with curated centroid keywords, so search is centroid-first: human-curated signal scores ahead of raw text. Five workflow recipes chain the pieces into repeatable routines — a morning brief, a deep research pull, an index refresh.

A local reading cockpit over six hundred research PDFs and the workspace's markdown: split reader, full-text PDF search lane, and a judgments overlay — stars, tiers, tags, notes, side-by-side pairs, thought-routing edges — kept in a store that joins to the corpus on content hash. Read-only against source files: the deck records judgment, never mutates the library.

An arXiv ingestion-and-synthesis pipeline: papers are enriched from public APIs with a local cache, scored against the workspace's research domains by a weighted keyword taxonomy plus domain-centroid embeddings, and emitted as dossier nodes that feed the knowledge graph. Includes an autonomous keyword-driven discovery loop and a daily orchestrator.

A local adaptation of Karpathy's MIT-licensed autoresearch harness — an autonomous loop that edits a small GPT's training script, runs fixed five-minute budgets, and keeps only changes that lower validation loss. The local contribution is a validation proof-gate: a sub-ten-second invariant check (model structure, shapes, loss finiteness, gradient flow, optimizer state) that must pass before any run is trusted. A real logged run exists: baseline 3.196 → 3.064 val bits-per-byte.

A multi-agent analysis dashboard (paper-trading research only — nothing here is financial advice) built around one idea: use Convergent Cross Mapping over Takens embeddings to separate true causality from correlation, with a deterministic Black–Scholes engine for verdicts and a language model confined to qualitative critique. Fourteen views in a brutalist observer aesthetic.

The narration layer of the research system: every research pull is logged, every session leaves a dispatch, and a brief writer turns raw pull results into narrated significance paragraphs tied to the active framework and prior briefs — a fourth memory layer over the library.

The portal's own semantic layer, honed to this corpus. At build time every document is chunked, embedded, and distilled into static geometry: embedding-nearest neighbors, nine-bucket category affinity, the five strongest lines per document by centrality, and cross-shelf candidate crossings feeding the Registry. Locally, a warm-once server answers free-text queries — ranked documents with distances plus the query's own affinity profile — with a terminal twin for the classifier-demo readout.

The implementation stratum behind several published threads — code, not documents. spectral/: laplacian, ricci_flow, spectral_field. holy_grail/: auto_retopo, conformal_quad_solver, gw_compatibility, island_detector, settling. hybrid_state/: state, sync — the State ≠ Mutation separation as running code. schwarz_christoffel/: biarc_boundary, solver — the conformal mapping lane.