what-counts-as-evidencelisted

# What counts as evidence **A signal is evidence only of what it actually observed.** Everything below is that sentence applied to the signals people habitually over-read. This skill owns what a signal proves. It does not choose which signals to gather, and it is guidance rather than a checklist — the table below is calibration, not an inventory to work through. This is the discipline that matters most when the work is done by an agent, because an agent will produce a confident summary either way. The summary is not the evidence. The summary is a claim *about* the evidence, and the two come apart silently. ## Signals that are not evidence Each of these is a real thing to know, and none of them proves what it is routinely taken to prove. | Signal | What it actually shows | What it does not show | |---|---|---| | Coverage | the line executed | that the scenario is correct | | A passing rerun | this run passed | that anything was fixed, or that the failure was infrastructure | | One timeout | this run exceeded a deadline | that there is a product race | | A fixed sleep | time passed | that setup completed or cleanup settled | | A stress run | it survived N repetitions | what a deterministic regression would show | | An exit code of zero | the command ran to completion | that it did the work you wanted | | A queued merge | the request was accepted | that it landed | | A structural check on a translated pair | the structure matches | that the meaning matches | | A zero-hit s