Earthlight is a program to train a foundation model on the planet's official record: every verified statistic Starwell collects, learned as one joint structure. Not a chatbot over data. A model of Earth's measurable state: how prices, people, housing, output, and trade move together across every country that publishes numbers.
Foundation models transformed text and images because vast, structured corpora existed to train on. Time-series and statistical modeling never had one: the world's numbers were scattered across hundreds of portals, unharmonized and unversioned. Researchers have pointed to exactly this gap as the reason the field lagged. Starwell closes it as a side effect of its normal operation: every observation it serves is stored with provenance, harmonized meaning, and full revision history. The data layer and the training corpus are the same asset. An idea this shape was proposed before, as Europe's Living Earth Simulator in 2011, and failed for lack of data, models, and compute. All three constraints have since reversed.
PHASE 0
Backtest today's open time-series foundation models on point-in-time nowcasts of Canadian and US indicators, using Starwell's revision store so no model ever sees the future. Published as the Earthlight-0 technical report.
PHASE 1
A transformer trained on the Starwell corpus itself, conditioned on what each series means, where, and in what units. It learns the joint structure of the measurable world: lead-lag relationships, cross-country propagation, seasonal and regime patterns. Ships as a nowcasting and anomaly-detection API on Starwell.
PHASE 2
The corpus holds something rare: every series is paired with the official prose that describes it, releases, notes, definitions. That pairing lets us teach a compact language model to read the state of the world numerically, the way vision-language models learned to read images. A small model that answers in words, grounded in learned numbers.
PHASE 3
Openly licensed satellite imagery and geospatial layers joined to the statistical spine. The precedent is a decade of economics: night lights alone have long been used to estimate economic activity where statistics are sparse. Earthlight extends the record to places the record misses.
Nowcasts of indicators weeks before official release, and estimates for countries whose statistics lag by quarters. Anomaly detection that flags releases inconsistent with the learned structure of the world, including statistics that may not add up. And economic fingerprints: embeddings of places that make questions like "find the region most like ours, ten years ago" a single query, for site selection, investment, and policy research.
Official statistics are small data by machine learning standards, so Earthlight promises calibrated prediction, never prophecy. Relationships learned under one policy regime can break when policy changes; economists call this the Lucas critique, and it is why Earthlight does not sell policy counterfactuals as causal truth. Every public claim ships with point-in-time backtests and calibration curves. A model of the world should be the first to admit what it cannot know.
Built on Starwell, the verified data layer for AI. Only openly licensed official sources enter the corpus, each with citation and attribution: likely the cleanest training corpus a foundation model has ever had.
Researchers: if you work on time-series foundation models, nowcasting, official statistics, or grounded language models, we want to compare notes. research@datastam.ai
Investors: Earthlight is long-horizon research built on revenue-generating infrastructure, with a training corpus that accumulates daily and cannot be backfilled by anyone who starts later. If you fund boundary-pushing work, write to invest@datastam.ai