An open data pipeline tracing the lineage, migration and divergence of research on the citation graph — seeded at Bell Labs' terahertz work (Auston, 1975). Start anywhere and Cladus keeps handing you the next thing.
Every paper you open teaches Cladus. It clusters what you keep returning to into mixes, and any paper can seed a station that keeps going.
The same graph that recommends things can also place you on it. Every topic you read gets two coordinates — how crowded it already is, and whether its share of new work is growing — and your reading is the weighted average.
Every paper, researcher and institution collected from the seed outward. Open one to follow its lineage — or start a station from it.
Founders and eras, crowded fields and untouched combinations, and the next wave as the model sees it. Every number is computed from the collected citation graph.
This project treats the evolution of science with the toolkit of population genetics and phylogeography. The metaphor is not decoration — it is the method: phylogeny estimation, migration flows, and founder-effect detection, applied directly to the citation graph. What follows is what that method actually returned on this corpus — the branches, the crowded ground, the blanks — followed by how it was built and what it cannot see.
collector/analysis/analysis/recsys.pyweb/| Year | Work | Role |
|---|---|---|
| 1975 | Picosecond photoconducting switch — D. H. Auston | Root of the lineage |
| 1984 | Picosecond photoconducting Hertzian dipoles | Origin of THz emission |
| 1988 | Subpicosecond photoconducting dipole antennas | The antenna line |
| 1988 | Electrooptic generation/detection of femtosecond pulses | The electro-optic line |
| 1990 | Generation of femtosecond EM pulses | Generalized fs pulses |
data/events.db, and “Reset taste” erases it. Coverage, name disambiguation and
language biases are documented as known limitations.The geometry above has a specific weakness, and it is worth naming before proposing a cure: it reads a paper as a bag of words, so it cannot tell a paper that measures a thing from one that reviews it. Both say “terahertz spectroscopy” and both land in the same neighbourhood. That is the gap a language model would close, and it suggests four jobs in the order they would pay off.
Claim extraction. Read each abstract into a structured triple — material, method, quantity measured — and cluster on those instead of on vocabulary. A research space would then be defined by what people do, which is what the word was supposed to mean.
Naming. A cluster currently called “conductivity · characterization” is a list of tokens the statistics liked. A model reading the twenty most central papers could give it the name a researcher in the field would use, with those twenty papers cited as the reason.
Bridge triage. The under-linked pairs above are candidates, and most candidates are coincidence — two areas can sit close in word-space and stay apart for an excellent physical reason. An agent with retrieval over both sides could sort the accidents from the real openings, and say which it is.
A question interface. Tool access to this same database — “who moved from photoconductive antennas into topological materials, and when” — answered by running queries, not by recalling training data.
Three constraints would apply to all four. Every generated sentence carries
the work ids it was derived from, and a sentence without ids does not render.
Generated text is stored in its own table, marked as generated, and never joined
into atlas.db without a human pass — a corpus that quietly mixes
collected and inferred records is no longer evidence. And the site stays fully
functional with the model switched off, which is also why it is off by default:
the figures above are reproducible by anyone holding the same database, and an
API-backed sentence is not.