METHODOLOGY

Every claim here either keeps its support attached or it gets cut.

This site is small enough to review by hand. That constraint drives everything else.

How essays get built

An essay starts because something broke, or a claim did not survive contact, or an experiment surprised me. Never from a keyword list.

Draft, then attack the draft. The test: does a reader leave with a sharper model than they walked in with, or just a more comfortable one?

Evidence

Source quality over source count. When a number is approximate, it says so. When I don't know, it says that too.

  • Direct observation
  • Primary documents
  • First-party data
  • Secondary commentary, last and labeled as such

THE WORKED EXAMPLE

A constant that was never measured

A file in my research vault stated a magnetization constant of 0.76 at specific parameters, formatted like a measurement. It was never measured.

It originated as a hypothetical target on line 290 of a working session. It failed to appear across the entire rest of that session. It got compressed into a reference document as a finding anyway. The result actually recorded at those parameters was 0.000390625.

I published the full trace, including the part where my own audit of it overreached, characterized the wrong model, and had to be corrected by a reviewer.

That is what the evidence standard means here. Not a promise to be careful. A documented case of catching myself, and of being caught.

AI usage

I use AI heavily: drafting, comparison, editing, code. I don't publish unattended output.

The failure mode I actually worry about isn't fabrication. It's compression. A long, messy exploration gets summarized. The summary keeps what's short and confident and drops what's long and equivocal. A tentative number arrives downstream as a finding with no trace of how tentative it was. Nobody lies at any step.

That happened to me. It is the case documented above.

So review here is specifically about that: checking whether a claim in a finished piece still has its support attached, or whether the support got summarized away.

Corrections

Errors get fixed on the page, with the correction visible rather than quietly patched over.

Outdated pages get updated, redirected, or removed. Nothing sits with a stale date on it pretending to be current.