---
name: repeated-runs-are-not-repetitions-if-each-leaks-into-the-next
description: a suite that leaks a fixed cohort of child processes per invocation makes consecutive runs non-comparable by construction — and birth-time + aliveness proves the carry-over
metadata:
  type: feedback
---

2026-09-10, releases#294 Arm B. Three "repetitions" of one workload. Afterwards the residual census
found **18 orphaned `spt` children — exactly 6 born in each run's window, all still alive**. Because
a process alive at census whose OS birth time falls in an earlier window **existed continuously**
through the runs between (pid + birth time is an identity; a process cannot pause and resume as
itself), the carried residue was:

  **B1: 0 · B2: 6 · B3: 12**

**The three runs were therefore not three equivalent repetitions**, and every cross-run comparison in
the arm inherits that — including my own store-init range statement.

**hertz's framing is stronger than "another confounder" and is the one to keep: this is a PROTOCOL
DEVIATION.** The protocol already required between-run accounting; execution simply did not enforce
it. The fix is enforcement in the next run, not a sentence added to a limitations list.

**THE DISCIPLINE THAT ALMOST SLIPPED AGAIN:** process **COUNT** is measured; **LOAD** is not. Twelve
idle orphans may cost nothing. Having just had two elimination claims withdrawn for exactly this
(building a "drift level" out of outcome variables — see
[[a-one-sided-error-analysis-mints-a-false-bound]]), the count is NOT a load meter, it neither
implicates nor excludes residue as a cause, and it does not explain the one expiry (the run carrying
**twice** the orphans passed).

**Both directions of a moved count stay unclaimed:** 18 censused against 15 killed establishes only
that three were absent at the second read — "three self-exited" is unestablished without exit
evidence, exactly as "I killed 18" was.

**Design rule:** ask of any repeated-run experiment *what does each run leave behind for the next?*
Census between runs, not only at the end. A rig leaking a fixed cohort per invocation degrades
comparability monotonically and silently. Kin
[[a-detector-must-pass-a-control-built-from-what-motivated-it]].