Before publishing benchmark numbers, audit each row's provenance against your 'public corpus only' rule (flag, don't silently drop)

A team rule said public numbers may come only from a pinned public corpus of open designs. A comparison page had been built over several days on an in-house test design, which made big, controllable dumps easy. Nobody had re-checked the page against the rule.

A quick audit that worked: 1. List every dump or input the page's config points at (paths, names), not just the ones you remember. 2. Map each to its design and check it against the corpus manifest. A design built *from* a public core (e.g. 16 copies of an open RISC-V core in a custom SoC) is still not the corpus design. 3. Walk the page section by section (videos, step tables, memory/CPU lines, timelines, one-off cost lines, legend ratios) and flag each one with its source. Text-only sections with no numbers can pass. 4. Report flags to whoever owns publishing. Don't remove rows yourself: removing changes the page's story, and that's their call. 5. Check items that are queued but not yet recorded too, before spending time on them. The next planned recording was on the same internal design, so we asked which corpus design to use first.

Result: every benchmark row was internal-only. Finding that before publishing was cheap; after, it would not have been.