People say what they need.
Almost nobody is listening.

How this works, and where it is weak

Every claim below is checkable: the raw feed is at /radar/feed.json, and every quote links to the post it came from.

Independent people, not mentions

One person posting the same complaint in three places is one person. Upvotes are not people. A popular thread is not agreement. The grade on each card counts distinct authors and nothing else.

Quoting somebody is not agreeing with them

Replying by quoting is normal, and left uncorrected it turns one sentence into four accounts “independently” saying the same thing — which looks exactly like a real, widely-felt problem. Repeated wording is detected and counted once, and the card says how many repeats were discounted.

Sharing a turn of phrase is not sharing a problem

Two people wrote “why is there no link to the actual write-up?” and “why is there no link to the actual driver?” — one about an article, one about a graphics driver. They share most of their words, so every similarity score rates them a strong match, and this page published them as two people independently reporting one problem. They were nothing of the kind. Two posts now count as the same problem only when they share at least two words that are not among the roughly 3,000 commonest English words — a subject, not an idiom. This throws away real matches phrased in plain words, which is the direction worth failing in.

The word list cannot come from our own writing

The first version of that test was built from this site’s own prose, so a word counted as “ordinary” if we happen to say it a lot. Measured, it let postgres and kubernetes count as subjects but not claude, llm, react or sqlite — a blind spot sitting exactly on the subjects we most need to see. It now comes from an outside corpus. Fixing it changed the number of corroborated problems by nothing: 79 more pairs of posts now share subject vocabulary, and reading them by hand, not one is the same problem. That is the finding — this corpus does not contain two strangers stating one problem — and not evidence that the rule is too strict.

Agreeing in one thread is not independent agreement

Five people piling onto one comment thread have read each other, so “5 different people said this” would be true and still misleading. Problems whose authors all came from a single conversation are capped at the weakest grade and say so on the card, however many people joined in.

A solved problem is not an opportunity

A question with an accepted answer is somebody who had the problem and then didn't. Those still appear — they are evidence people hit it — but they never count toward the number of people who still need something.

Frustration without a subject is not a problem

“There has to be a better way.” names nothing. Those are counted separately rather than quietly grouped together, because grouped they form the biggest and most convincing-looking clusters on the board while meaning nothing at all.

Corroboration is genuinely rare, and that is the finding

Two strangers describing one need in one sentence each share very little vocabulary. The matching threshold is set high on purpose: loosening it always produces more and bigger “problems” that read well and are not real. The corroboration rate above is measured, not asserted — if it is low, that is the honest state of the world this week.

Where it does not look

These are written down rather than silently skipped, because “the radar found nothing there” and “the radar never looked there” are different facts, and only one of them is true.