Reproducibility
Most of this can be checked without me. Some of it cannot, and I would rather tell you which is which than let you find out on your own time.
What you can check today
The mathematics. Every derivation is in the papers and the papers are at the DOI. If a step does not follow, it does not follow, and it makes no difference at all whether you trust the person who wrote it.
The files. Every artefact in the December bundle has a published SHA-256. Recompute one and compare. It takes a single command and nobody’s permission. Do it in your browser if you prefer.
The instrument. In July my evaluator was scoring measurement failure as a perfect result. An empty panel came back as zero misalignment, which is the best score available, so a run that measured nothing reported success in every arm. I found it on 2 August and fixed it that day. The attack that caught it is now a permanent test, and you can run it yourself: it is indexed, with its command, in the repository’s experiments index.
What you cannot check yet
I should be clear about what I am claiming.
The empirical work is unregistered and unreplicated. The stability criterion is proved inside a stated model and has not been tested on a real one. The April verification is a result supplied to me, which I have not reproduced locally.
Until confirmatory work is registered and run, treat every empirical number on this site as provisional. That includes the numbers that favour me. Especially those.
What would prove this wrong
Every paper states the condition that would kill it, and states it before the data exists rather than after. That is the only version of falsifiability worth anything.