Every record in the catalog plotted on two axes from the in-catalog
citation graph. The X axis counts inbound cites edges (academic influence — how many other records
cite this one); the Y axis counts inbound built-on + integrates-with edges (commercial
adoption — how many other records build on or integrate against this
one). The chart is sliced into four named quadrants by the
non-zero medians on each axis. Hover any point for a third lens — inbound runtime-dependency edges (issue
#44) — answering "who breaks if this record goes down". Headline
finding: Academic citation and commercial integration are essentially orthogonal at the current corpus stage: the "Both" corner stays sparse even as either single-axis quadrant fills. Papers and production code are not yet on the same lineage.
Bridge surprises
Records whose betweenness rank beats their raw inbound rank by the largest margin — they sit on more shortest paths between
clusters than their popularity suggests. Often the
mid-tier orchestration / spec records that hold the graph together
without being widely cited.
Recent method papers — theorized, no distinct product
+119
0.147
0
2
+ 5 more positive
surprises tracked internally.
Nucleus (k-core = 3)
The deepest k-core in the graph: every record in this set has at
least 3 neighbours that are themselves
in the set. Empirically the zone where next products / papers
originate — the densest mutually-connected substrate of the
corpus.
60 records in the nucleus.
K-core distribution across the full graph: 057011482134360
Systems above the median on both axes — academia cites them and engineers integrate against them. These are the rare nodes that have crossed from research result into production substrate.
So what. Treat this as the watch list. If your system lives here, it has durability signal from two independent communities; protect the API surface and resist breaking changes. If a system you depend on is here, the bus factor is probably fine.
High inbound integration but low (or zero) inbound citations. Working code that downstream teams build on — but that hasn't been picked up by the literature, often because it predates the current research wave or because it's industry-led infrastructure.
So what. If your system is here, the open question is whether academic study should catch up — papers citing real production systems are more reproducible than papers citing only other papers. Worth seeding a benchmark / case study.
High citation count, low integration count. The paper is in the air but nothing in the catalog has been built on top of it yet. That's a normal phase for recent work — adoption lags citation by 12–24 months — but persistent orphans suggest a result that doesn't translate to a usable artefact.
So what. Track these for adoption signals: a research orphan that picks up its first `built-on` edge is the strongest leading indicator we have for a paper crossing over. If your system is here, ship a reference implementation.
Records with zero or below-median inbound edges on both axes. Most of the corpus lives here. That isn't a failure — most recent work simply has not had time to accumulate graph centrality, and many entries are products with no academic reach by design.
So what. Don't over-read absence. Re-check this quadrant after the next two extraction rounds — anything that moves out of the tail is a candidate for closer review.
Cutoffs: cites > 2, integrations > 1.5 (non-zero medians on each axis, recomputed against
the current filter).
Section colour key (top 10)
Recent method papers — theorized, no distinct product (191)
Use-case-specific agent harnesses (87)
Vertical / domain-specific AI memory (64)
Training infrastructure (51)
Dedicated memory layers (41)
Agent frameworks (no first-party memory layer) (39)
Retrieval-as-memory hybrids (37)
Agent IDEs & coding harnesses (32)
Memory benchmarks & evaluation (32)
Framework-embedded memory (31)
Marker size encodes tier: T1 largest (7px), T5 smallest (3px).
Faded markers are records with zero inbound edges of either type
(the bulk of the long tail). Translucent orange hexes are the
density underlay — darker = more records binned into that cell.
How centrality is computed (issue #46)
The structural-centrality measures above run on the undirected projection of the in-catalog edge graph — every record pair joined by any edge of
any type counts as adjacent. Centrality is a property of the WHOLE
graph, so the bridge and nucleus callouts do not recompute when you toggle the tier/section filters above
(those filters only affect the scatter projection).
Betweenness centrality
For each pair of records (s, t), the fraction of shortest paths between them that pass through a given record v. Records with high
betweenness sit on many shortest paths — they "bridge" otherwise
disconnected clusters. Implemented via Brandes' algorithm
(O(V·E), single-source BFS + back-accumulation). Scores are
normalised to [0, 1] by dividing through the max betweenness in the
graph; the legend value is therefore a relative measure, not the
textbook combinatorial invariant. Brandes, U. (1986). "A faster algorithm for betweenness
centrality." J. Math. Sociol. 25 (2): 163–177.
K-core decomposition
The k-core of a graph is the maximal subgraph in which every node has
degree ≥ k within that subgraph. The "coreness" of a node
is the largest k such that the node belongs to a k-core. Computed
by the standard peeling algorithm (Batagelj & Zaversnik 2003,
O(E)). The highest k-core is the nucleus — the densest
mutually-connected substrate of the corpus, empirically where the
next round of products / papers most often originates. Seidman, S. (1983). "Network structure and minimum degree." Social Networks 5 (3): 269–287. · Batagelj, V. &
Zaversnik, M. (2003). "An O(m) algorithm for cores decomposition
of networks." arXiv:cs/0310049.
Bridge surprise (Δ rank)
Rank-by-betweenness minus rank-by-raw-inbound (lower rank number =
better, so the difference is positive when betweenness rank is
better than inbound rank). A positive value means the record sits
on more shortest paths than its inbound count suggests — it earns
its structural position from bridging, not from being
widely cited. This is the metric the "Bridge surprises" callout
above is sorted by. We require non-zero betweenness to qualify
(otherwise records with no signal at all would clog the top).
Computed at SvelteKit prerender time over 912 records and 514 unique undirected adjacencies. Maximum k-core observed: 3. Nucleus size: 60 records.