---
name: load-knob-must-be-proven-to-bite
description: "When you ADD load to hunt a rare face, measure a span that must stretch under it — otherwise zero reds is a fact about your knob, not about the tree."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 65448c94-60f8-4bce-88b7-3548e33546a0
  modified: 2026-08-21T10:20:34.018Z
---

Hunting releases#199's 1-in-20 face (2026-08-21) I escalated to 4-, 10- and
12-way concurrency and got 610 green bring-ups. **That number means nothing on
its own.** "I added load and it stayed green" and "I did not actually add load"
produce identical evidence, and the second is the comfortable reading.

**Why:** the escalation is an instrument like any other, and an instrument whose
effect nobody measured is untested by construction. I only earned the right to
report the greens because I pulled the ConPTY-create span out of the phase stamps
and watched it scale monotonically with the knob: **9-16 ms quiet, 42 ms at
4-way, 117 ms at 10-way, 294-326 ms at 12-way.** That is a 20-30x stretch, so the
knob demonstrably bit — *and* it was still ~100x short of the 30s bound, which is
itself the finding: this box cannot reach the window at this scale, and saying so
is a measurement rather than a shrug.

**How to apply:**
1. **Pick the span BEFORE the run.** Name the thing that must stretch if the load
   is real (a create, a lock wait, a queue depth), and make sure the instrument
   already reports it. Retrofitting the span after a green run invites picking
   one that flatters the story.
2. **Report the scaling curve beside the red count**, always. A rate with no
   knob-effect number is an anecdote about a box.
3. **Compare the stretched tail against the BOUND you are trying to cross.** If
   the tail is orders of magnitude short, further escalation of the same kind is
   predictable waste — change the mechanism or stop, and say which.
4. A single outlier in a stretched tail is **one sample**. Mine (1427 ms in a
   span whose median is ~1 ms) did not recur in 240 further observations; it goes
   in the report as an unreproduced observation with its location, never as a
   candidate.

Inverse of [[load-claim-needs-measured-axes]] (never *assert* load as a cause
unmeasured); this is never *credit yourself* for load you did not verify you
applied. Rate arithmetic still needs [[base-rate-is-over-cell-executions-not-suite-runs]]
— and check [[filed-rate-may-predate-your-base]] before trusting the target rate
at all. Kin: [[non-vacuity-check-can-fail-for-the-defects-reason]].