Lab Notes

David Lindgreen — Lab Notes

Real data, inspectable systems, honest limits — across AI, physics, astrophysics, evolution, biology, chemistry, climate, neuroscience, and data observability.

Explore the researchRead the articles

What this is

A growing collection of interactive research projects and real findings, written up in plain language. Every project begins with a bounded, falsifiable question and a protocol appropriate to its design. Blinded or preregistered studies are identified as such; known-result reproductions are labelled explicitly. Every report includes what was actually found — including the parts that didn't confirm cleanly. Nothing here claims to be peer-reviewed novel science. What it claims is narrower and checkable: real public data, a documented method, and an honest account of where that method holds up and where it doesn't.

Why this exists

Most of what looks like "using AI to do research" online is a demo dressed up as a discovery. The goal here is the opposite: hand you something you can actually poke at — an interactive tool, not just a chart — a written explanation of what it found and why that's interesting, and a straight line to the actual public data and cited papers if you want to go further than we did. If any of these articles makes you want to open the tool, pull the real dataset yourself, or read the paper we're testing against, that's the point. Everything here is MIT-licensed and reproducible — clone any of these repos and check our work.

Explore the evidence

Research questions. Inspectable answers.

Filter the portfolio, choose a study, and inspect its question, evidence, result, and boundary before opening the full laboratory.

Machine learning / representation geometry

Neural Geometry Lab

Question
After zero training error, does neural-collapse geometry keep improving—and track unseen-writer accuracy?
Evidence
30 frozen MLP runs, three stress conditions, and an official writer-disjoint digit split
Finding
Two of four gates passed; noisy training had slightly better median NC2 but much worse accuracy and NC1.
Boundary
One small MLP and dataset; seeds quantify algorithmic sensitivity, and no single collapse coordinate certifies generalization.
Read the articleOpen the interactive lab

Academic evidence index

Trace every claim to evidence—and its limit.

“Strength” here is always scoped: confidence in the narrow tested claim, never a universal score across disciplines.

Astrophysics

FRB Atlas

Evidence design
Single-catalog reanalysis with a disclosed post-hoc source-level check
Evidence assessment
Moderate for this catalog's sample-dependent contrast; weak for distinct FRB populations or other surveys.

Supporting or limiting evidence

DatasetVizieR J/ApJS/257/59 table 2InterpretationFull research reportReproduceFrozen result registry

A reusable method

How to read a scientific result

Each lab uses the same four-part discipline. Use it here, then carry it into any paper, chart, model, or headline you encounter.

  1. Ask a bounded question

    Replace a broad topic with a claim that data could genuinely contradict.

  2. Inspect the evidence

    Check where the data came from, what was excluded, and what was measured.

  3. Read uncertainty first

    Look at intervals, sample structure, robustness checks, and null results.

  4. Stop at the boundary

    A result supports only the population, method, and conditions actually tested.

The articles

Explore further