Standards on the desk — a field note on slow reading

This is a short one. A field note, not an analysis.

Tuesday mornings, I read the standards. Not papers. Not newsletters. Not industry blogs. The actual standards.

A printout of IEC 61508 — the functional-safety standard the older safety-critical industries have been organised around for forty years. A printout of Annex III of the EU Machinery Regulation 2023/1230, the regulatory text that applies in January 2027. A printout of ISO 13849, the machinery-specific performance-level standard. An espresso. A pen. No laptop.

Why not papers

The academic literature in robotics is good. It moves fast and surfaces a lot of novel work. It is also, for the purposes of understanding the safety gap, not the right entry point.

The standards are the written-down version of forty years of arguments about what actually prevents the worst outcomes in systems that can hurt people. They were written by people who had to defend every sentence in front of a working group, an industry, and — eventually — a regulator. A paper proposes. A standard commits.

The reason the safety gap in Physical AI keeps getting misframed in trade press is that most commentary on it comes from people who have read the papers and maybe the press releases, but not the standards. The standards are where the real constraints live.

Why not newsletters

Newsletters compress. Their business model requires it. That's fine for following the industry — I read several and I'll cite them when they're useful — but it's not where you go to understand a regulation that's two years from applying to every machine you place on the EU market.

You go to the regulatory text.

Why slow

The standards are not difficult reading in the sense that a dense mathematics paper is difficult. They are difficult in the sense that they reward re-reading. A clause that seems obvious on the first pass turns out, on the third pass, to constrain an architectural choice in a way that matters. A definition that you skipped over in section 3 turns out to be the hinge of an entire compliance argument in section 8.

You can't speed-read a standard and have it produce the understanding that slow reading produces. I've tried. It doesn't work.

What this looks like in practice

Two hours. One topic. One printout. A cup of coffee. No notifications. Notes in the margin in the same pen every week, because I find uniform handwriting across weeks helps me remember what I thought three Tuesdays ago.

The analysis I publish here is downstream of that habit. When I can be specific about a regulatory date or an architectural requirement, it's because I've read the paragraph it comes from. When I can't be specific, I'm careful to flag it.

There's nothing glamorous about this practice. Nobody posts about it on LinkedIn. That's most of the reason it's an underrated skill.

@misc{melchior2026standards,
  author       = {Melchior, Mati},
  title        = {{Standards on the desk — a field note on slow reading}},
  howpublished = {Physical AI Safety},
  year         = {2026},
  month        = apr,
  date         = {2026-04-22},
  url          = {https://physical-ai-safety.com/blog/standards-on-the-desk}
}