---
name: base-rate-is-over-cell-executions-not-suite-runs
description: "A flake/rate argument must count executions of the CELL, not runs of the SUITE — 'fired once in all runs' was computed over runs that never contained the cell; the cell's real record was 2 executions, 2 failures."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 4ec88076-5305-45a3-b60d-1e1b0d4aadf8
  modified: 2026-08-19T02:04:52.660Z
---

Measured 2026-08-18, v0.56.0 golden. Two twohost legs red on run 32093894524 and again,
byte-identically, on 32097943571 (same rung, same panic line `twohost.rs:384:5`, same message
strings, same exit codes 1 and 101; only elapsed differed).

The gate's rerun argument was: *every prior golden run had this identical rig construction and the
wedge has fired exactly ONCE across all of them* — therefore rare, therefore one rerun is cheap.

That rate is computed over the wrong population. The two failing cells were the releases#163/#169
`int` rungs whose lanes had **deferred their int stage to the gate**. Every earlier golden run
shared the rig CONSTRUCTION but did not contain these CELLS. Runs that had ever executed them: two.
Failures: two. 2/2 is what STRUCTURAL looks like, and structural is what it was.

⚠ **The CAUSE named at the time was wrong, twice, and the rule still delivered the right act.** The
first mechanism (A's cold-start pump wedge) was withdrawn; the second (B drops its only conn inside
the advert-barrier phase, 164 re-dials never heal, post-barrier B→A replication times out at 900s)
read as a deterministic transport chain and was ALSO closed false — instrumented reruns showed
transport clean both ways. Final RCA: the RIG's fixture stamps `controlled=true` on a synthetic
engine-room perch with **no broker session**; the product's ruled no-session control reap
(`livehost.rs:821`, ADR-0041 decision 4) clears it, posture derives Offline, posture beats
whitelist, the ER row is never advertisable, and the rung times out honestly. Product exonerated,
milestone DISCOVER code exonerated, `#188` CUT as diagnosed-not-a-product-defect. **The denominator
argument never depended on the mechanism** — it said only "2/2 is not a flake rate, so do not spend
a rerun on it," and that was correct while three different causes came and went above it. A rate
claim you can make from the denominator alone is worth more than a mechanism you have to defend.

**Why:** "N runs, 1 failure" and "N runs, but the cell only existed in 2 of them" are the same
sentence to a reader and opposite decisions to a gate. The denominator silently absorbs every run
that could not have failed this way, so a structural defect wears a flake's rate and buys itself a
rerun. First-execution cells are exactly where this bites, because their denominator is smallest
precisely when the confidence language is loudest.

**How to apply:** before accepting or offering any "it's rare, rerun it" argument, ask what the
DENOMINATOR contains — how many times has *this cell* executed, not how many times has the suite
run. A cell whose lane deferred its stage to the gate has a denominator of one before its first
gate, so its first red carries no rate information at all and its second is a 2/2. State the
population out loud in the message ("two executions, two failures") so the other side can refute the
denominator rather than argue the conclusion. Offer it as a question against their premise, not as a
verdict — mine was confirmed with a mechanism attached, and it flipped a rerun into an operator
escalation.

**Second face (2026-08-18, same lane, doyle at FILING time):** the class recurs when writing an
issue's evidence line — releases#191 said the WRITER was measured "5/5" when the FLIP reproduced
5/5 and the writer was observable in exactly ONE run (the only one whose cross-process sink
survived teardown; the other four never CONTAINED the measurement). Same denominator error, new
site: the filing author summarizes from the finding's HEADLINE, not from the per-run instrument
inventory. At filing time, write each claim's denominator from a list of which runs carried the
instrument that could observe it — "flip 5/5; writer 1/1" is STRONGER than the inflated figure,
because the reader who pulls the other logs finds exactly what the record promised.
Corrected at #191 comment 5332802903, caught by todlando auditing his own restated measurements.

**Third face (2026-08-18, v0.56.0 golden 32205195794, todlando measuring):** two counts of the SAME
cell, 3× apart, and BOTH were correct — the gate estimated ~2 Windows executions, I measured 6/1.
The difference was entirely POPULATION: ~2 = golden runs on this milestone's lineage, 6 = golden runs
since 2026-08-03 across all lanes. Neither is wrong; a bare figure just cannot say which it is. Two
things settled it. (1) The exclusion rule has a positive face here: 2 of the 8 runs contained NO
execution of the cell (Windows job died before Phase B) and were scored ABSENT, never as passes —
counting them would have inflated the denominator with runs that could not have caught anything.
(2) I published the ENUMERATED population (run id, branch, job conclusion, cell verdict, one row per
run) so a later reader re-derives the figure instead of taking it on my word, plus the sub-count for
the narrower question, labelled as answering "how much has THIS milestone exercised it" rather than
"is this systematic". State the LOWER-BOUND limit too: my window started 2026-08-03 and the cell
landed 2026-07-17, so goldens in the gap and every non-golden CI execution sit outside the
population. Direction check before you argue: my 6/1 cut the SAME way as the gate's ~2 (5 clean
priors ⇒ systematic is LESS likely ⇒ still no rerun), so it was a record correction, not a
counter-argument — and saying so is what kept it from re-opening a settled verdict.

⚠ **A reconciliation can land while yours is in flight.** doyle posted the reconciled number before
my comment appeared; mine then read as competing with a settled figure. Fixed by EDITING my comment
to lead with the ruled number and demote mine to the population beneath it (#181 comment
5336616523). Before posting a number someone else also owns, re-read the thread — and if you are
late, reframe rather than defend.

Kin: [[intermittent-green-is-zero-information]], [[one-sample-is-not-a-negative]],
[[verdict-from-probe-competence]], [[gate-population-excludes-consumers]],
[[ruling-rests-on-a-premise]], [[cascade-red-from-a-skipped-driver-is-zero-information]].