Skip to content

Using the map

The cartography is a single self-contained page — open it here — with five tabs along the top. This page explains what each tab shows and how to read it. Before drawing conclusions, read the headline caveats: origins are panel-conditional, the corpus is 1995 left-censored, and the trajectory layer is a descriptive projection, not a forecast.

Every panel is built from the derived tables documented in the Data dictionary; each tab below links to the relevant section there.

1. Topic landscape

A scatter of every leaf topic (149 of them) — one point per topic, sized by how many corpus papers it contains and coloured by its origin journal's diffusion role (source / bridge / terminal). This is the base map: it shows the shape of the topic space and which topics were first observed (panel-conditionally) in source-type vs terminal-type journals.

  • Read it as: a bird's-eye view of what the panel is about and where each topic entered the panel.
  • Watch out: colour reflects the origin journal's role, and origin is panel-conditional and left-censored — a topic coloured as "source-origin" was first observed in this panel in a source-type journal, not necessarily first in the world.
  • Backing tables: topic_hierarchy.parquet, cascade/origin_attribution.parquet, roles/role_scores.parquet (dictionary).

2. Cascade flows

How topics move between journals. Three views:

  • a lead–lag heatmap of inter-journal time offsets (who tends to enter a shared topic before whom);
  • a directed seeding network (arrows from seeding journals to following journals);

  • an adoption-by-breadth curve (how a topic's reach across journals grows over time).

  • Read it as: the diffusion structure of the panel — the evidence behind the "cascade is real" claim (journal roles are stable under jackknife, rank-correlation ≈ 0.9993).

  • Watch out: lead–lag and seeding are onset-timing signals and so are sensitive to the 1995 left-censoring (caveat 2).
  • Backing tables: cascade/lag_matrix.parquet, flow/flow_leaf.parquet, cascade/diffusion_curves.parquet (dictionary).

3. Journal roles

Where each of the 78 journals sits in the diffusion network. Two bar views:

  • source / bridge / terminal component bars (the z-scored sub-scores that decide each journal's role);

  • citational-velocity bars (how fast each journal's papers accrue citations).

  • Read it as: the journal-level summary of the cascade. This is where the headline finding is visible: the generalist source layer splits — PNAS is the #1 source, while NEJM is the most terminal and Lancet / JAMA / Nature / Science are also terminal sinks.

  • Watch out: roles are within-panel; the seeding sub-signal is left-censoring-sensitive while flow/betweenness are robust. A journal labelled source is a source relative to this panel.
  • Backing tables: roles/role_scores.parquet, roles/velocity.parquet (dictionary).

4. Novelty origins

Where novelty comes from. Three views:

  • sector novelty — novelty by institution sector (this signal is near-null across sectors);
  • a geography ranking — novelty by country;

  • top recombinant topics — the topics whose papers most combine references from across distinct topics (combinatorial novelty).

  • Read it as: a descriptive breakdown of novelty by who and where; the recombination list highlights the most combinatorially novel topic areas.

  • Watch out: the sector signal is near-null — do not over-read small differences. Country rows flagged low-sample should be filtered out for headline reads. All origin signals are panel-conditional and left-censored.
  • Backing tables: origins/sector_novelty.parquet, origins/geo_novelty.parquet, origins/recombination_by_topic.parquet (dictionary).

5. Trajectories

The descriptive "future" layer. For each topic, the model fits the observed 1995–2025 series of its panel share (and raw volume) and extrapolates +5 years to 2030 with an 80% band. The tab has:

  • a direction distribution — across all 149 topics the split is flat 81 / falling 38 / rising 30;
  • top rising and top falling movers drawn as fan charts (the projected point with its widening 80% band);
  • a per-topic dropdown selector (a Plotly updatemenus control): pick any one of the 149 topics and the chart redraws to its observed + projected share with the 80% band. Use this to inspect a specific topic rather than only the headline movers.

!!! warning "Trajectories are descriptive, not a forecast" This layer is a descriptive state-space projection (local-linear-trend on logit(share) / log(volume)), extrapolated forward. It is not causal, not a validated predictive model, and is explicitly not the V1 F3 emergence GNN — that predictive model was signed NULL at Gate G4. The 80% bands are extrapolation uncertainty, not a tested forecast. share is panel-conditional. See caveat 4.

  • Backing tables: trajectory/trajectory_summary.parquet, trajectory/trajectory_series.parquet (dictionary).

General reading tips

  • Hover for the underlying numbers (topic labels, journal slugs, lags, shares); the map is interactive Plotly.
  • Join key. Almost everything is keyed on topic_id (0–148), resolved to a human-readable label via topic_hierarchy.parquet.
  • Grain. The headline cartography is the leaf grain (149 topics). Some tables also carry a coarser mid grain; the map shows leaf.
  • When in doubt about what a field means, the Data dictionary is authoritative; to rebuild the map from source, see Reproduce.