For each of the 8 detected lineages,
project where the next entry is likely to land based on the cadence of
past releases. Exploratory — watch list, not prediction.
Read this first. Most lineages on this page have only
3–5 dated members. A cadence computed from 2 or 3 inter-arrival gaps
is informative directionally (is this family active or
stalled?) but is not a statistical prediction. The "next expected"
quarter is a watch trigger — a date to check the source
feeds — not a forecast to defend. If a family dries up, gets absorbed,
or accelerates by a single release, the projection moves by a year.
How the cadence is computed
For each lineage we list its members in descending order by created date, parsed to quarter granularity
(year × 4 + quarter).
The cadence is the mean of the inter-arrival
gaps in quarters, with a floor of 0.5 quarters per gap so that
same-quarter clusters don't crash the average to zero.
The prediction interval is the population
standard deviation of those gaps. We surface it as
"Q3 2026 ± 2 quarters" — read this as "if the cadence holds,
the next entry will probably land within this window."
Discount the projection when (a) the lineage
has fewer than 5 members (small-sample badge), (b) the cadence
is built from fewer than 3 gaps (insufficient-data badge —
we hide the date entirely), or (c) the gap spread is large
relative to the mean.
So what? A short-cadence lineage means the architectural idea is still
earning new contributions — reading the next 2–3 arXiv updates in
that lineage is high-leverage. A long-cadence lineage means the idea
is either mature (settled, low marginal returns) or stalled
(researchers moved on). Adjacent lineages are where cross-pollination shows up first: work in one family
routinely informs predictions for its neighbours.
8 of 8 lineages shown ·
0 small-sample (N<5) ·
0 insufficient data (cadence intervals <3)
if the cadence holds, next expected ≈ 2026-Q3 ± 1 quarter
Leading edge (3 most-recent dated members)
Themes the leading edge is exploring
claude
cloud
consolidation
fewer
files
opaque
preprint
prevents
Adjacent lineages (cross-pollination candidates)
What to watch for (editorial)
Watch for cross-tool standardisation. 30+ implementations of the same "text file the model reads at session start" pattern cannot continue without convergence on a shared format.
if the cadence holds, next expected ≈ 2026-Q1 ± 1 quarter
Leading edge (3 most-recent dated members)
Themes the leading edge is exploring
dynamic
graphrag
memory
preprint
raptor
aaai
accommodation
agentic
Adjacent lineages (cross-pollination candidates)
What to watch for (editorial)
Watch for hybrid retrieval at scale. Most current papers stay below 10M entities; the breakout will be the first credible Graph-RAG result on a >100M-entity corpus.
if the cadence holds, next expected ≈ 2025-Q4 ± 6 quarters
Leading edge (3 most-recent dated members)
Themes the leading edge is exploring
agents
atari-100k
environment
interaction
limited
preprint
world-model
agent
Adjacent lineages (cross-pollination candidates)
What to watch for (editorial)
Watch for the first commercial spinoff. DeepMind has the densest academic cluster here but no T1 product has shipped yet — when one does, it will validate the whole branch.