Verdict: declining — 3 events in 2026 against 3 in 2025; 264 of the 267
better-objective events were posted 2015–2021
CVRPLIB X-instance record events
Click a point to open the update page it was posted on.
Definition
The capacitated vehicle-routing problem asks for minimum-cost routes that
serve customers without exceeding vehicle capacity. CVRPLIB accepts improved
solutions, checks them, and posts each change to its chronological Updates
ledger.
This series freezes the 100 X instances introduced as one designed cohort. A
"discovery" is either a newly posted lower objective for an X instance or a
later proof that the standing objective is optimal. The two are counted
separately: finding a better route and proving no better route exists are
different algorithmic events. Every row is dated by the public ledger posting
date; receipt, paper and posting dates are not mixed.
Facts
events: 289 event rows: 267 better-objective events and 22
optimality-proof events over 2015–2026
last pre-2026 objective: the last X-objective change before 2026 was
posted 2021-06-30
2026 posting: three objective rows in the 2026-07-04 posting: one for
X-n979-k58 and two successive values for X-n1001-k43
proofs: the 22 optimality-proof events cover 22 distinct instances
The collection-wide cumulative index redraws this
cohort as instances remaining without an optimality proof:
Method
fetch.py walks all five pages of the Updates ledger, selects
instance names matching the X convention, normalizes two historical omissions
of the n, and writes a row for each objective improvement or proof phrase.
Combined "improved and proven optimal" announcements intentionally create two
rows for an instance. The 2026 announcement reports receipt dates inside a
July 4 posting; the data uses the public ledger date consistently for every
row.
figure.py counts rows by posting year and stacks proofs above
objective changes without treating them as the same event type.
check.py recomputes the fact lines above from the CSV.
Limitations
the ledger starts late. The update ledger begins after the X instances
were introduced; it is not a reconstruction of their pre-ledger initial
solutions.
batch postings. A posting can contain work performed over a longer
interval, so annual counts are publication cadence as well as discovery
cadence.
the parser classifies the page's words. It does not independently
verify that every "improved" value beats the previous value, although
CVRPLIB states that it checks submissions.
the frozen cohort excludes later sets. Freezing X avoids
benchmark-composition drift but excludes later Loggi, ORTEC, XML and XL
records, including the 2026 BKS Challenge.
unweighted counts. Event counts do not weight the magnitude or
difficulty of improvements.
AI attribution
No AI system or language model is identified in the update text for this
fixed cohort in the entries vendored through 2026-07-04. The 2026 entries are
attributed to named optimization researchers. This is an authorship statement
about the ledger's text, not a claim that no AI component was used anywhere
inside a solver.
Sources
Updates ledger
— the chronological posting log every row is fetched from; it states that
submitted improvements are checked and that an optimality claim requires a
citable method.
