A model is nerfed when it feels worse than it did last week. Shorter answers, more refusals, weaker reasoning, work it used to get right, and no word from the maker that anything changed.
Sometimes that is real. A quantised deployment, a changed system prompt, a routing change. Sometimes it is the tool (harness) around the model, or a harder prompt than you remember writing, or memory fuckn with you.
How the number works
Votes are anonymous and gated by Turnstile. Each model takes one vote per person per hour. A nerfed vote has to name what got worse: code, instructions, refusals or speed. Voting fine needs no reason.
The share on a card is the last 24 hours, updated as votes land. Once a day it gets written down as a fixed point, and those points draw the 30-day strip.
What it cannot tell you
A spike is not proof a model changed. People who are annoyed turn up to vote, people having a normal day mostly do not, and one popular thread can move a line on its own. We run no benchmarks and we cannot see anyone's weights, serving config or routing.
What it is good for is timing. If one model's share jumps and its neighbours stay flat, that is worth asking about.