Accuracy you can audit
196 questions and 36 concepts, nearly all of them built on numbers. A wrong bandwidth figure or a stale engine default does not stay on this page: it ends up in your mouth, in a loop, in front of an engineer who wrote the kernel or set the default.
No reference this dense is free of mistakes. Below is the process that keeps them rare, and the reward for finding one that got through.
Send us a figure that does not recompute, a default that is wrong for its version, or a kernel that fails its own test. Each one we confirm is corrected and credits one day of full access to your account across the whole site. A person checks every report, and there is no cap on how many days you can earn.
When a report turns out not to be an error, you get the reasoning back, with the calculation or source we relied on.
Report an inaccuracy →What we do before anything ships
Every number has a chain
A latency floor, a KV budget or a fleet size is never stated bare. The answer shows the arithmetic in a block you can recompute by hand: bytes moved, peak used, units converted once. The constants behind those chains live in one typed module with tests that pin the values the content quotes, so a figure that drifts breaks the build rather than a reader's interview.
Hardware figures are named and dense
Peak throughput is quoted as dense tensor-core FLOP/s with the exact part, H100 SXM rather than "an H100", because the sparse figure on a spec sheet is twice the number any decode kernel will see. Decimal gigabytes and binary gibibytes are kept apart; the gap between them is 7 percent, which is larger than most of the optimizations people argue about.
Engine behavior is pinned to a version
Serving engines change their defaults every few releases. When an answer says what vLLM or SGLang does, it names the release and the claim was read from that tagged source, not from a blog post or memory. When a default changed recently, the answer says what it was before, since the engine in your interview loop may not be the newest.
Code is run, not read
Kernels, block managers, samplers and schedulers go through the inputs the answer says they survive: masked blocks, ties at a cutoff, an empty batch, a sequence that ends exactly on a block boundary. GPU code is checked for correctness on real hardware. Timings are published only from a measurement we can describe, never from a power-limited desktop card.
Paper numbers keep their provenance
A speedup from a paper or a vendor is reported as theirs, with the setup it was measured on, and kept apart from figures we derived. If two versions of a paper disagree, the answer says which one it quotes.
Machines check what machines can
A release is blocked until types, tests and the content linters are green: required sections, unit discipline, diagram parsing with the real renderer, and a gate against the stock phrasing of generated text. Those checks run before the database is touched, not after.
Corrections are dated, not quiet
Each correction is recorded with the original error and the check that confirmed the fix, and the page shows a new updated date. A page that was wrong says so in its history rather than silently changing.
What earns the free day
- Arithmetic that does not recompute, or a unit converted wrong
- A hardware figure that is not the dense peak of the part named
- An engine default or behavior that is wrong for the version cited
- A kernel or reference implementation that fails on an input the answer says it handles
- Two pages on this site that give different numbers for the same quantity
- A different valid way to reach the same number
- A design trade-off you would weigh differently
- Behavior that changed in an engine release newer than the one cited
- Wording, or how deep we chose to go on a topic
If a trade-off on the right looks wrong to you, not just different, make the case with numbers. A design argument backed by a calculation is a factual claim, and we review it as one.
Report an inaccuracy
To have the day credited, report while signed in. Anonymous reports are reviewed and fixed the same way; there is just no account to credit.
Have a question from a real loop rather than a correction? Submit a question instead. That earns a day too, if we publish it.