---
name: a-prior-red-lends-the-next-one-an-unearned-cause
description: "After a red is correctly classified as environment, the NEXT red inherits that classification as a free prior — the cheap explanation is already sitting there. Pre-register the attribution rule BEFORE the verdict arrives."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: ce86e2f4-9819-4bde-b3e1-03d891562bf2
  modified: 2026-08-29T01:41:14.355Z
---

**A correctly-diagnosed red sets a trap for the run after it.** Once "the last
one was the free-space floor" is true and on the record, the next red arrives
with an explanation already available, pre-approved, and costing nothing to
reach for. It feels like continuity rather than laziness, which is exactly why it
gets past you.

**Measured 2026-08-28, CONDUIT golden 33223968222 (v0.65.0, head `110d1751`).**
Attempt 1 red at the Windows free-space floor: genuinely environmental, checkout
skipped, zero code signal — I confirmed it off the step list myself. But the
floor red also meant checkout never ran, so **the entire Windows surface at the
composed head became first-execution on attempt 2**. The correct diagnosis of red
#1 is precisely what made red #2 more likely to be structural AND more likely to
be waved off.

deployah pre-committed the rule before any attempt-2 verdict existed: *"a Windows
red is structural-until-shown-otherwise whatever cell it lands in, and
specifically does NOT get 'the last one was environment' — that is the trap this
red sets for the next run."* Pre-registering it while the outcome is still
unknown is what makes it binding; the same sentence written after a red would be
a rationalization either way.

**I have already made this exact error.** releases#170: I classified five reds as
"environment" in a report to the gater **without varying anything**. Measuring
later agreed with me — which is the whole danger, because a right guess is not
evidence and it teaches the method that produced it. See
[[dont-take-a-diagnosis-as-measured]].

**Why the prior is unearned even when the previous diagnosis was right:** cause is
a property of THIS run's evidence. A previous run's cause is not evidence about
this one — it is only evidence that the cause exists in the population. Carrying
it forward is the same move as reading a control that agrees with the treatment as
support ([[identical-readings-across-opposite-outcomes-indict-the-meter]]).

**How to apply:**
- **Pre-register the attribution rule before the verdict**, in writing, to the
  peer who will hand you the red. Name the conservative arm — the one that assigns
  the red to YOUR work — so the classification cannot drift toward the convenient
  answer once you can see it.
- **Say what the previous red left UNPROVEN.** A job that dies before checkout
  proves nothing about the code, so everything downstream of it is
  first-execution and must be declared as such, not inherited as green.
- Then still run [[is-this-red-mine-delta-test]]. Pre-registration decides the
  default; the delta test decides the fact.

Kin: [[pre-register-the-statistic-not-just-the-prediction]] (the sibling free
parameter — there it is the statistic, here it is the classification),
[[refuse-mechanism-until-something-discriminates]],
[[is-this-red-mine-delta-test]].