---
name: one-sample-is-not-a-negative
description: "A single passing run is not a non-reproduction, and a probe can pass for a reason unrelated to what it claims — get a distribution on both sides, and probe the handle you hold."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: bfe75fb9-b545-4df4-90d8-cde9a051e778
  modified: 2026-08-02T15:02:35.239Z
---

Two self-inflicted false negatives in one session (2026-08-02, BAROMETER golden triage):

1. **n=1 reported as a non-reproduction.** A golden red ran once at the buggy base on the platform
   it failed on, passed, and I reported "it does not reproduce — that arm reopens". It reproduces
   **4/15 idle, 7/15 under load**. The failure was nondeterministic, so one sample carried no
   information at all. Fix: for any red not proven deterministic, run a **distribution at the base
   AND at the fix on the same rig** (15+15 was enough here: 4/15 & 7/15 vs 0/30).
2. **A probe that passed for the wrong reason.** Checking whether a kill landed, I asked a
   pid-EXISTENCE probe about a process the test still held an open `Child` handle to — an already
   killed process still read alive, so the probe passed and would have reported the kill as
   refused. `try_wait()` on the handle is the honest read. **A liveness probe against a process you
   hold a handle to is not a liveness probe.**

**Why:** both shapes produce a green that means nothing, and both are indistinguishable from the
real result they impersonate — the same class as a zero-match filter reading as absent.

**How to apply:** before reporting an absence, ask what a broken detector would look like here and
whether it would look exactly like this. Then either name the mechanism that makes the negative
sound, or take samples until the rate is a number.

Related: [[instrument-soundness-guards]], [[zero-match-filter-reads-as-absent]],
[[discriminator-question]], [[premature-closure-guards]].