Five publishable headlines from the catalog — 1,005 records × 69 columns across 38 sections, 2,542 typed edges. Each card cites the analytical view that produced it; full derivations live in analysis.md.
1 of 100 priority-cohort agent products
Inside the agent stack, only LangChain's SemanticCache ships a true semantic-cache layer in the priority cohort. The capability exists, but it is being sold by the data and gateway layer (for example Redis LangCache), not by agent frameworks or model providers.
922 of 1,005 products · only 4 scores independently verified
The product × benchmark matrix is 123 × 25 with 173 filled cells — the story is the empty cells. Of the 173 filled, the integrity split is 115 peer-reviewed / 4 independently-verified / 27 vendor-claimed / 27 disputed. Almost every claim is a paper or a vendor blog.
Inbound runtime-deps: Claude 62 · GPT 52 · MCP spec 34 · Gemini 22 · Qwen 16
From the T2-1 runtime-dependency graph (212 edges of 'X depends on Y at runtime'). A protocol specification ranking #3 — ahead of every foundation model except Claude and GPT — is the finding that wasn't visible in citation-graph hubs. The MCP spec is becoming a substrate, not just a protocol.
0 inbound edges · 0.71 normalised betweenness — highest non-trivial in the graph
Sits on the shortest path between the MCP-spec substrate cluster and the Zep / Graphiti citation-anchor pair. A high-betweenness / low-inbound node is the structural definition of an under-acknowledged-but-load-bearing connector. Top-5 bridge-surprise: Graphiti MCP Server, MAGMA, Memformers, MemEvolve, RGMem.
108 of 140 FM-dependent rows depend on OpenAI / Anthropic / Google
Of the 140 rows (~16% of non-FM rows) that name a foundation model, 77.1% sit on the OpenAI / Anthropic / Google single-vendor-risk tier. The long tail (Qwen 16, Mistral 12, Cohere 8, DeepSeek 6, Jamba 3, Grok 2, Nova 2) is sparse. The substrate-risk map is concentrated on three labs.