BDS Pipeline Performance Notes
Observed behavior across indexing, extraction, QA, and entity-card generation workloads.
Coming soonResearch Blog
Lab notes from operational AI research.
Benchmark notes, retrieval findings, hardware experiments, pipeline behavior, and visual research updates.
Reference · Start here
The Nuance Glossary
Every signal, column, and concept these lab notes rely on — Channel, Similarity, the Result Critic, anchors, wikilink repair, the corpus-shape trap — defined once, in plain language. The posts link back to it.
Open the Glossary →Methodology · 2026-08-22
When to Reach for GraphRAG — and When Not To
GraphRAG is the most over-sold idea in retrieval. A precision-first, verifier-gated relationship graph across land-use, litigation, and safety corpora — and the honest conditions under which it's worth building at all.
Read →Lab Notes · 2026-06-29
Facts in Code, Judgment in the Model
Building an OSINT synthesis layer taught us two things the hard way — don't let a language model reason over a noisy proximity field, and the most valuable thing it can tell you is “don't bother.”
Read →Methodology · 2026-06-27
Building a Test Corpus Is the Hard Part
Why realistic inference data is harder to assemble than the pipeline is to tune — two public-source systems, the difficulties, and the limits we haven’t closed.
Read →Lab Notes · 2026-06-07
Corpus-Shape Trap
Why “not in the corpus” is a useful finding, not a system failure.
Read →Agentic Systems · 2026-05-27
Switching Agents at Phase Two
How a VIOSINT build moved from Hermes/Gemma4 to Claude Opus, why the pivot happened, and what agent orchestration judgment is worth.
Read →Meaning vs Word Match
Why vector, BM25, and both signals need different interpretation rules.
Coming soon