Every discovery has an address.
We already know how to read it.
Singularity is a forecasting engine for science — reading 163 million works across every discipline, tested on fifteen years of history, confirmed twice with rules sealed in advance, and already producing research. It names, with a deadline and a strength measured on the past, where the next advances will appear. This page is a technical sheet: what it is, how it works, the evidence and what has come out of it so far. We are not selling anything. If you want to support the work, with funds, time or ideas, or simply learn more, write to us.
The thesis in one paragraph.
The world spends more than two trillion dollars a year on research and development, and it decides where to look largely by expert intuition. Yet the history of science shows a regular, learnable pattern: breakthroughs are often old methods arriving in new fields — and those arrivals leave measurable traces years in advance. Whoever reads those traces first knows where progress will happen next.
- What it is
- Forecasts of where each method will appear next, for every field of science, with a deadline and a probability.
- Useful to
- Researchers and students choosing what to work on; labs and companies scouting technology; funders and governments placing grants; investors spotting the next deep-tech wave.
- Evidence
- Fifteen years of backtests on all of science, two preregistered confirmations, public registries of timestamped forecasts, and studies already started from our own forecasts.
- Where we are
- Three model generations validated. The flagship, v3, reads all of science (163 million works, 252 subfields) and passed two sealed confirmations; 146 of its forecasts are fingerprinted and open.
How it works — without the recipe.
Five stages, from public data to a verified forecast. We tell you exactly what goes in, what comes out and how it is tested. The signals and models in the middle are our intellectual property.
Ingest
Titles, abstracts, fields, dates and authorship links of the scholarly record. Open metadata only (arXiv, OpenAlex — CC0).
Recognise
A curated vocabulary of scientific concepts finds the methods in each text, merging synonyms and acronyms.
Map
A year-by-year map of which field uses which method, how fields resemble each other, and who works across them.
Learn
An ensemble of models learns, from decades of past crossings, which early signals came before a method entered a field.
Forecast & verify
Every possible field × method crossing is ranked; the top ones are published with a probability, a date and a fingerprint, and scored when the deadline comes.
- The data sources, all open
- The test protocol, written before the test
- Every forecast, with date and fingerprint
- Every outcome, failures included
- The performance, against simple baselines
- The exact signals and how they are computed
- The vocabulary and its entity resolution
- The models, their training and calibration
- The full rankings beyond what is shown here
- Our pipeline of discoveries in progress
Why you can trust the numbers.
Forecasting the past is easy if you peek at the answers. These are the rules we set ourselves so that we cannot.
Never sees the future
Each backtest year is forecast by a model trained only on what was known then — and only on outcomes already complete at that date.
Judged on the hard cases
We report accuracy separately on untouched crossings — fields with zero papers using the method — where trend-following cannot help.
One shot, preregistered
Before declaring a new model better, we write the test down, then run it once on years never used. v3 passed twice: 58 of its top 100 untouched crossings opened on 2017–2022, and all four views passed on 2023.
Averages of the six sealed confirmation years. Only crossings into subfields that had no work with the method at forecast time. The obvious forecast is strong in the AI era — which is why we also show the views without AI. Year-by-year detail on the track record.
Forecasts that already turned into research.
We did not stop at predicting. We took our own forecasts and turned them into results — each study took days. This is the discovery studio, working at the smallest scale.
Why now. Why us.
The record of science became open.
Since 2022 the metadata of almost every scholarly work is free to use. What used to require expensive licences is now a public good.
Trying a method became cheap.
With modern computing and AI, testing an idea in a new field takes days, not years. The bottleneck has moved from answering questions to choosing them.
Nobody else keeps score.
Trend reports describe what is already hot. Very few publish specific, dated forecasts and let them be judged.
A track record that cannot be copied backwards.
Every forecast carries a date and a fingerprint. A competitor can copy a model; nobody can go back and register forecasts in 2026.
Compounding data.
Each year adds a new set of outcomes to learn from and a new set of forecasts to judge. The engine improves with time.
Built with very little.
Three generations were built and tested with minimal resources, and the first studies came out of them. More resources would mean more of science read more deeply, more forecasts tested and more studies.
If you want to help, write to us.
Funds, time, data, a lab, an idea, a criticism: every kind of help is welcome, and we will work out together how. We have nothing to sell and no fixed plan; we want to make progress, and to get better at it along the way.