Crash Safely
This page explains how to read the site: which parts are written by people, which part is drafted by a machine, and what neither of them is allowed to do. It exists because a site whose headlines are jokes owes its readers a clear statement about where the jokes stop.
The jokes are in the titles
"My GPU Filed for Remote Work" is a joke about renting accelerators. The article under it is a straight piece about what changes when your compute stops being local — what you actually trade away, and which of your habits stop working. The headline is the packaging.
Nothing on this site describes an event that did not happen. There are no invented outages, no fictional postmortems, and no anonymised war stories that are really composites. When a failure mode is described here, it is described as a pattern — a thing that happens to systems of a certain shape — and not as an incident with a date that somebody lived through.
No numbers of our own
This site runs no production infrastructure and measures nothing. That has one firm consequence: no figure here is presented as our own measurement.
You will not find a latency number, an availability percentage, a mean time to recovery, a throughput figure or a recommended timeout value stated in this site's own voice. Where such a number appears, it comes from a published source and is attributed to that source in the same sentence, so you can go and check it. This matters more than it might seem, because readers copy numbers out of articles into configuration files, and a plausible invented threshold is a genuinely harmful thing to publish.
The same applies to security advice. Explaining why something is dangerous is part of the job. Recommending it is not.
The automated section
Blast Radius is written differently from the rest of the site, and its pages say so at the top rather than leaving you to work it out.
Each piece starts from a topic chosen by a human and written down in advance — a specific claim about how systems fail that is worth examining because it is wrong, incomplete, or true only under conditions nobody states. A pipeline fetches that topic's sources, which are verified article URLs rather than domains, drafts an analysis, and puts the draft through a quality gate.
The gate rejects work mechanically, on properties that can actually be checked: whether the sources were genuinely read, whether any long passage is lifted verbatim, whether the piece engages the claim it is supposed to be about, whether it states a measurement without attributing it, whether it reads as a product recommendation, and whether it claims first-hand experience of running systems the pipeline has never run. A draft that fails is redrafted with the reasons fed back, up to three times. If all three fail, the topic is retired and nothing is published that day. An empty slot is an acceptable outcome; filler is not.
What survives becomes a merge request that sits for a review window before it goes live. That review is the actual quality control. The gate is a floor — it stops specific known-bad outcomes reaching the site — and it cannot tell you whether an argument is any good. Treating a passing check as an endorsement would be the whole mistake.
What the gate cannot do
It cannot verify that a claim is true. It cannot tell whether the reasoning holds. It checks shape, not substance, and it is published here as a floor rather than advertised as a guarantee.
If something in that section is wrong, it got past both a machine that was not looking for wrongness and a human who missed it. Corrections are welcome through the chat widget on any page.
Tracking
The site uses Google Analytics to count visits, and a Chatwoot chat widget that loads when the page loads. There is no advertising, no affiliate link, and nothing for sale, so nothing here is trying to move you toward a purchase.