Open research program · computing signal out of public data · updated July 2026

The signal is already in the data. Nobody went and got it.

Across many fields, rich public records sit un-computed. The discipline that owns the data does not use the method, and the discipline that owns the method does not know the data is there. We go get the signal. We compute the ocean’s unobserved past from physics, reconstruct lost economic history from ledgers that had to balance, and re-test published scientific claims on their own data. Different fields, one instinct, checkable every time, honest throughout, nulls included.

Computing what was never measuredEconomic history from ledgers that must balanceRe-testing published claims on their own dataBuilt-in checks, or it does not countOpen data, honest nulls

Where the work has gone

Three angles, one method

Computational ecology · the worked example

The eel spawning ground is “warming in place”

Where does the European eel spawn, and is climate moving it? The 22 to 24°C spawning isotherms slide about 190 km poleward, faster than the SST-gradient front that bounds the spawning zone. The simple front-displacement story is wrong. We couple differentiable models to the physics of currents, temperature, and the Earth’s magnetic field to compute what no instrument ever recorded. Our strongest paper-ready result.

Explore →

The honest null

A cross-species law that did not survive independent data

The eel pattern suggested a general anchoring theory of climate and reproduction. It blind-tested at about 88% across 101 species, then collapsed when we hid the species identity, and returned a null on independent fish-recruitment data. We publish the failure in full. A hypothesis, not a law.

Explore →

Marine carbon removal · new

We measured where ocean carbon removal actually goes

Ocean alkalinity enhancement is credited on models too coarse to see the fronts where treated surface water is pulled down before it can trade its carbon with the air. We measured that missing process in two independent high-resolution ocean models. At open-ocean fronts, 16 to 38 percent of the water is subducted below the winter mixed layer first, and coarsened to crediting resolution the number drops to about zero. The carbon is not lost, it re-emerges over decades, outside the horizon it was sold on. The over-count concentrates on the subpolar and Southern Ocean waters where alkalinity is most often proposed. The full portfolio, science, policy, and investment, is drafted and open for review, carrying the registry gap of about 0.5 percent of contracted tonnes physically realized and the investability screen. Honest throughout, including the sign that flips for biological methods.

Explore →

Computational history · new

Three colonial ledgers that check themselves

The data-side twin of the eel work. One computes the past from physics, this one recovers it from records that had to balance. Dutch East India Company, Batavia: 18,321 voyages, the books reconcile to 98.6 percent. Sound Toll Registers: 2.15 million ship passages through the Øresund, 1497 to 1857, where the Amsterdam to St-Petersburg and London trade axis flips between 1807 and 1815, confirmed by the Napoleonic closure of the strait that dropped traffic 98 percent in two years. Spanish-empire royal treasuries: 202,585 line items, where barely one percent of inter-treasury transfers appear in both the sending and receiving books, so the missing half is the finding.

Explore →

Reproducibility · new

Re-testing published claims on their own data

Many published machine-learning models report near-perfect accuracy that comes from data leakage, not real signal. On fully public ICU data, a single correct change to how the data is split moves a model from 0.93 to 0.80. We are building an open screener that flags leakage-prone setups from a paper and its public data. We reproduce the authors’ own number first, we critique methods not motives, and we report risk, not verdicts. A public good, not an accusation.

Method · working prototype

The method

How we choose what to compute

Rich public data the field already trusts, a question that field cares about, a mature method that maps onto it, and above all a way to check the answer. Built-in ground truth, like a ledger that must balance or a fragment that must physically fit, or a hard external one, like a modern instrument re-measuring the same thing. If we cannot check it, we do not claim it. The hardest part is novelty, finding the rich record the owning field has not yet computed.

On-demand engine

Open data asset

A harmonized 14k-species climate-reproduction database

Reproductive thermal limits, thermal tolerance, range-shift outcomes, and life-history traits from six public databases, GBIF-resolved, with an autonomous ingestion pipeline. A reusable community resource, independent of any model.

Explore →

The full notebook

Read the whole research notebook

Timeline, datasets, every test we ran, what held up and what did not. Leading thoughts at the top, full progression below. Everything traces to a number. Raw data is reproducible from DOIs.

Explore →

Provenance

Every data source we reviewed

The exhaustive catalogue: what we ingested and analyzed, what is a standing ocean/climate input, what is request-gated and pending, and what we reviewed and set aside (with reasons). Licences and status throughout, nulls and embargoes included.

Explore →

For fun · sci-comm

10 Clickbait Articles After Researching Eel Sex

The whole saga, told the way the internet would tell it. An AI navigated by quantum mechanics, found a universal law, and was gloriously wrong. Every headline is clickbait, and every one is a real result.

Explore →

New · quantum

Can a quantum computer compute the compass in a bird’s eye?

The radical-pair magnetic compass may be how birds, turtles, and eels navigate. We ran it on IBM’s quantum hardware and mapped, honestly, where a quantum computer helps: a laptop still wins the tractable version, the field’s standard shortcut is ~2x off, raw device noise erases the signal, and the real quantum machine for the full problem is ~end of decade. We measured the limit instead of hyping it.

Explore →

New · open data platform

SorbentBase: an open database of water-treatment sorbent performance

Gas adsorption has NIST ISODB. Aqueous water treatment had nothing. We mined 26,000+ sorbent-performance measurements from the full text of ~1,700 open-access papers, with experimental conditions, mechanism, molar-normalized capacity, and a source link on every row. Queryable and live.

Explore →

Work in progress

Manuscripts & releases

Submesoscale subduction biases ocean-alkalinity-enhancement efficiency atlases at ocean fronts

Manuscript + full portfolio + open code · target Nature journal, seeking co-authors

Draft, in review

A Sargasso eel-spawning isotherm migration that outruns the front

arXiv preprint · climate / marine ecology

Draft, ready to finalize

A differentiable larval-transport engine that reconstructs spawn origins

arXiv preprint · methods, the engine behind the reconstructions

Draft

Why cross-species climate-vulnerability indices fail (the anchoring framework, retired)

arXiv preprint · methods / negative result

Drafted; framework retired after independent tests failed

A both-ends reconstruction of the Dutch East India Company’s intra-Asian trade

Deposited on Zenodo (10.5281/zenodo.21071241) · arXiv (econ.GN) next

Deposited, drafting

The Sound Toll network and the 1807 to 1815 reordering of northern trade

Deposited on Zenodo (10.5281/zenodo.21071241) · arXiv (econ.GN) next

Deposited, drafting

Reconciliation residuals as an anomaly engine in the Spanish-empire treasuries

Deposited on Zenodo (10.5281/zenodo.21071241) · arXiv (econ.GN) next

Deposited, drafting

An open leakage screener for published prediction models

arXiv preprint + Zenodo · open tool

Prototype

Open release: the harmonized climate-reproduction database

Zenodo dataset + data paper

Asset built, packaging

Under research · to consider

Open questions we’re still chasing

Open & reproducible

How we work

Every figure and claim links to a finding note and a script. Raw datasets are reproducible from their DOIs. We pick problems where the answer can be checked, by a ledger that must balance, a fragment that must fit, or an independent instrument, and we report only what the evidence supports, stating plainly what is not claimed. When a result fails, we publish the failure. We welcome data, critique, and collaboration.

We default to fast, open dissemination. Most results go to arXiv as preprints, and code and data are deposited to Zenodo with a citable DOI. The three history reconstructions above are deposited at doi.org/10.5281/zenodo.21071241. Formal peer review is worth the effort for some results and not for others, and we say which is which rather than letting work sit in a queue for a year.

Steps Ventures · an independent open-science effort.