2  The two datasets

The book runs on two public graphs, each used where it is strongest.

• Air Routes — Neptune’s own teaching dataset: ~3.4k airports, a route graph (airport↔︎airport) and a contains hierarchy (continent→airport). Numeric and categorical features, thin text. It carries the tiers 01–02 on-ramp.
• ogbn-arxiv — ~169k CS papers, ~1.16M citation edges, 40 subject classes, rich title+abstract text. It carries the tiers 03–04 spine, and is the dataset the graph-conformal literature reports on, so tier-04 is directly comparable to published results.

Both are fetched on demand and checksum-gated: a silently reissued source fails loudly rather than drifting the book’s numbers. Only the committed subset ID lists (data/ids/) ship in the repo — subset identity is committed, not reproduced from a seed.

The book’s teaching spine is connect(target) parity: a recipe is written once and the only thing that changes between a laptop and a GPU server is the target. Here we open the embedded engine on CPU with a file:// target; the keystone’s heavy emit opens a GPU server with connect("grpc://…") and runs the same verbs.

import tempfile

import jammi
from jammi_cookbook import datasets

db = jammi.connect(f"file://{tempfile.mkdtemp()}")
air = datasets.load_air_routes(db)
arxiv = datasets.load_ogbn_arxiv(db)

2.1 Air Routes

A registered source is queried as <source>.public.<source>.

def count(src):
    return db.sql(f"SELECT COUNT(*) AS n FROM {src}.public.{src}").to_pylist()[0]["n"]

print(f"airports:       {count(air.airports_source)}")
print(f"route edges:    {count(air.route_edges_source)}")
print(f"contains edges: {count(air.contains_edges_source)}")

# Each airport's continent is resolved from the continent→airport hierarchy.
atl = db.sql(
    f"SELECT code, city, country, continent "
    f"FROM {air.airports_source}.public.{air.airports_source} WHERE code = 'ATL'"
).to_pylist()
print("ATL:", atl[0])

airports:       3504
route edges:    50637
contains edges: 7008
ATL: {'code': 'ATL', 'city': 'Atlanta', 'country': 'US', 'continent': 'NA'}

2.2 ogbn-arxiv

The committed subset is a connected induced citation subgraph — a connected ball preserves the citation density and label homophily the tier-04 coverage crux depends on.

papers = arxiv.papers_source
print(f"papers (subset):     {count(papers)}")
print(f"citation edges:      {count(arxiv.cite_edges_source)}")
print("time-split sizes:    ", {k: len(v) for k, v in arxiv.split.items()})

top = db.sql(
    f"SELECT subject, COUNT(*) AS c FROM {papers}.public.{papers} "
    f"GROUP BY subject ORDER BY c DESC LIMIT 3"
).to_pylist()
print("top subjects:", [(r["subject"], r["c"]) for r in top])

papers (subset):     4000
citation edges:      9595
time-split sizes:     {'train': 804, 'valid': 1081, 'test': 2115}
top subjects: [('arxiv cs cv', 1491), ('arxiv cs lg', 1112), ('arxiv cs cl', 750)]

The citation graph is homophilous on the subject label — adjacent papers tend to share a subject far more than chance. That correlation, together with the dataset’s time-split, is exactly what breaks conformal exchangeability in tier 04 — where marginal subject-classification conformal under-covers and the textbook weighted-conformal remedy turns out to be a no-op (the shift is orthogonal to the score).

edges = db.sql(
    f"SELECT e.src AS s, e.dst AS d, a.subject AS sa, b.subject AS sb "
    f"FROM {arxiv.cite_edges_source}.public.{arxiv.cite_edges_source} e "
    f"JOIN {papers}.public.{papers} a ON e.src = a.paper_id "
    f"JOIN {papers}.public.{papers} b ON e.dst = b.paper_id"
).to_pylist()
homophily = sum(r["sa"] == r["sb"] for r in edges) / len(edges)
print(f"citation-edge subject homophily: {homophily:.3f}  (random ≈ {1 / 40:.3f})")

citation-edge subject homophily: 0.500  (random ≈ 0.025)