Six analytical views on top of the catalog substrate. Each surfaces a pattern that's hard to see in the main table — influence dynamics, survivorship, benchmark coverage, taxonomy recipes, vocabulary drift, forward projection.
Every system plotted on two axes: inbound citations (academic influence) × inbound integrations (commercial adoption). Surfaces research orphans, engineering wins, and the diagonal of both-cited-and-adopted. Now includes Brandes betweenness + k-core decomposition — the "bridge surprises" callout flags records that hold the graph together without showing up in raw citation counts, and the nucleus callout lists the densest mutually-connected substrate.
Active / Stale (>12mo) / Abandoned (>24mo). Cross-references latest-release and code-release against created date. Filters downstream analysis to alive systems without losing the dead-but-influential rows.
For each headline memory benchmark (LongMemEval, LoCoMo, BABILong, ConvoMem, RULER, MemoryBench), which systems published a score? Reveals which benchmarks have critical mass.
Companion to coverage. Classifies every benchmark mention as peer-reviewed / independently-verified / vendor-claimed / disputed / unverifiable, exposes gaming patterns, and surfaces the canonical disputed case (Mem0 vs Zep on LoCoMo).
Pivot of the benchmark coverage view, flipped from benchmark-centric to product-centric. Rows are products, columns are benchmarks, cells are coloured by integrity tier. The "refuses to publish" filter surfaces every product that mentions benchmarks but has zero peer-reviewed citations behind them.
Pivots the integrity classifications into a single composite trust score per benchmark. Sortable leaderboard with top-5 trusted / most-gamed callouts and per-benchmark drilldowns showing which systems contributed to each integrity bucket.
The #1 demand-signal question in agent infrastructure (807 HN hits, 89% LangChain-survey adoption, 6/7 published surveys). Matrix of products × observability tools (LangSmith, OpenTelemetry, Datadog, Helicone, Weave, Langfuse, Arize). Top ~100 priority rows backfilled; the rest surfaced as "unknown" so the coverage gap is explicit.
The #2 demand-signal question (171 HN mentions last 12mo, Sequoia framing on inference-cost sustainability, Datadog 2026 found 69% of input tokens are system-prompt overhead). Matrix of products × cost-control features (token budget, prompt caching, semantic caching, batch API, model routing, streaming-only, cost attribution). Per-row governance score (0/7 → 7/7). Top ~100 priority rows backfilled.
The next frontier after observability. LangChain State of Agent Engineering 2025: 89% have observability adopted but only 52% have evals — a 37-point structural gap. This view is the first catalog-level surface for it: side-by-side obs vs eval adoption %, plus an "eval orphan" filter that surfaces products with observability ≥1 tool but zero eval tools. Matrix of products × eval tools (LangSmith Evals, Braintrust, W&B Agent Eval, Helicone Evals, Custom test harness, Human-in-loop, Prod traffic replay). Top ~100 priority rows backfilled (same set as observability + cost-economics so cross-cuts work cleanly).
Clusters systems by their 7-axis taxonomy fingerprint. Surfaces recurring recipes (vector+semantic+persistent; graph+structural+temporal; file+text+manual; etc.) richer than any single axis.
Quarterly occurrence of key memory terms (episodic, working, parametric, agentic, lifelong, world-model) in descriptions and claims. Reveals what is emerging vs declining at the concept level.
For each detected lineage, what does the leading edge tell us about where the next paper/product is likely coming from? Pairs with the influence-vs-adoption view to anticipate Q3 additions.
Combines funding cadence, release recency, GH stars, mindshare/citations, and lineage membership into one velocity story per record. Six panels: cohort, substrate-dependency risk, consolidation candidates, billion-$ candidates, dying candidates, and a BIMATEM-style logistic S-curve fit per row that reads pre-growth / growth / saturation / decline straight off the inflection date — the math the Gartner Hype Cycle waves at without ever publishing.
Bibliometric staple since Henry Small 1973: cluster systems by how the community pairs them (co-citation) and by what they collectively cite (bibliographic coupling). Force-directed map plus a disagreement panel that flags pairs the community pairs tightly together but our hand-built taxonomy splits across sections — the headline finding no comparable catalog publishes.
Fits the Wang-Song-Barabási (Science 2013) log-normal citation model to every paper row with a citation trajectory — the academic gold-standard for paper-level impact prediction. Three views: papers most likely to break out (still-growing fastest), hyped-but-won't-last identification (saturation below cohort median), and sleeping beauties (slow-burn impact with low immediacy but long tail). The catalog's watchlist for what to add next quarter.