---
name: process-census-count-is-noise
description: "A process census's COUNT is not a measurement — short-lived CLI polls make it fluctuate wildly; only the classification by executable path is trendable."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: af542e3e-b09f-4831-8947-e03c0c37926b
  modified: 2026-08-04T07:02:48.352Z
---

Counting processes on a live fleet box and trending that number against an earlier record
manufactures leaks that are not there. Short-lived CLI invocations churn constantly, so any two
samples disagree.

Measured 2026-08-04 on HFENDULEAM, pre-fire census for golden attempt 5. Three `spt.exe` samples
inside ~90 seconds: **23, then 17, then 15**. Nothing leaked and nothing drained — the churn was
`spt endpoint digest <id> --json` invocations, other agents polling their inboxes, alive a second or
two apiece. The pre-attempt-4 record said "11 remaining fleet"; anyone diffing 23 against 11 reads a
doubling that never happened.

What IS trendable is the classification: every process ran the installed fleet binary
`C:\Users\decid\AppData\Local\spt-core\bin\spt.exe`, with ZERO under a worktree / target / temp /
runner / scratch / tests path. That answers the question a pre-fire census actually asks (are there
leaked test-path residents), and it is stable across samples because the durable set is stable —
one `daemon run --detached` plus its brain child, and the per-agent `rc <id>` / `api ... listen`
listeners.

The zero was positive-controlled before being believed: the same regex run against a synthetic
`.worktrees/…/target/debug/spt.exe`, a `Temp` path and an `actions-runner` path all matched, and the
fleet bin path did not. Absent, not unexpressible.

**Why:** a census is a claim about a POPULATION, and a raw count silently mixes a stable population
with a transient one. doyle logged this as the IR-14 count-disagreement lesson generalized from
artifact counts to process censuses.

**How to apply:** report a census as a classification with its filter and its positive control, never
as a bare number; if a count must appear, sample it more than once and say the range. When someone
hands you a count that disagrees with a prior record, ask what the population was before treating the
delta as a leak.

Kin: [[a-predicate-without-its-tool-is-not-evidence]] (root+dialect+tool+population),
[[zero-match-filter-reads-as-absent]], [[verdict-from-probe-competence]],
[[one-sample-is-not-a-negative]], [[precise-claim-is-falsifiable-vague-is-useless]].