---
name: first-execution-is-a-property-of-the-cell
description: A zero-CI-denominator cell stays zero-denominator on every platform — naming the box that failed last time narrows the pre-declaration to the wrong surface.
metadata:
  type: feedback
---

**First-execution is a property of the CELL, not of the box that failed last time.** I narrowed it
twice in one night and was wrong both times, in the same direction.

1. At CONDUIT intake I pre-declared ONE zero-denominator cell (`io_events_poll_e2e`). todlando
   pointed out the golden's Windows jobs had died at the free-space floor with **checkout skipped**,
   so the WHOLE Windows surface at the composed head was first-execution — not one cell.
2. Correcting that, I then pre-declared "the composed head's WINDOWS surface is a combination
   nothing has run." The respin golden went red on **LINUX**; Windows passed all four new cells.
   The cells were zero-denominator on BOTH platforms. I aimed at Windows because the earlier floor
   red had made Windows salient — i.e. I let the last failure pick the surface.

**Why:** a first-execution cell's CI denominator is zero everywhere until it runs everywhere. The
box that bit you last time is the most available prior and has no bearing on which box has never
executed the cell. Reaching for it feels like continuity — the same trap as reusing a correctly
diagnosed red as the prior for the next one.

**How to apply:** derive the first-execution set from the CELLS added or never-run at this head,
then state it per-platform-agnostic ("these N cells, on every axis"). Only narrow to a box when you
have measured that the other boxes already ran them. Say it BEFORE the run — a pre-declaration is
evidence only when written against an unknown outcome, and it is textually identical to a
rule chosen to fit the result, so the clock and the open question are what carry the weight.

Related: [[dont-take-a-diagnosis-as-measured]] ·
[[a-frontier-names-where-progress-stopped-not-the-cause]] · [[v0650-arc-in-flight]]