Independent researcher · Spain
I run small research programmes and try to break my own results before anyone else can.

A six-month quantitative research programme on crypto perpetual-futures microstructure. It asked whether a slow participant could extract systematic edge, and it could not. So I published the other thing it produced: a register of every claim the programme made that later turned out to be false, with the person or instrument that caught each one named.

git clone https://github.com/Lostmanu/ninety-three-wrong-claims
cd ninety-three-wrong-claims
python tools/recuento_auditoria.py --check

That command parses the register's own tables and prints the count. Nobody types the number: a tool regenerates it, and refuses to run if two entries share an id. That control exists because the count once said three different figures at once in the same document.

The programme's single positive estimate, about +18 basis points, met a review protocol that had been frozen months earlier. All eight of its lenses returned a threat that applies and the verdict was destroyed. The estimate was retired. That is in the repository too, with the dates.

→ Read the register, the paper and the tool

Those are the tools the work above actually used: a data collector running unattended on a rented Linux server, an analysis tree in Python with 476 tests, and a continuous-integration job that runs six checks in 72 seconds and fails closed.

• Preregistration before data. What counts as success is written and committed before the numbers exist.
• Controls that prove they bite. Every guard has a row that breaks one line of source and requires a named test to fail. If no test fails, the guard is decorative and the harness says so.
• The number is not typed by the person reporting it. A tool emits it with the command that produced it.
• The register. Every claim of mine that turned out to be false gets written down, before it is fixed, with who caught it. Including the ones nobody would ever have known about.

I work with AI systems, which write code and audit it. Most of the entries in that register are theirs. That is the point of keeping it.

_{Everything here is unfunded, unaffiliated and not peer reviewed. The register is the honest part.}