---
name: an-unexercised-prediction-is-not-a-correct-one
description: "A pre-registered prediction whose triggering condition never arises is UNEXERCISED, not correct — log it that way or it sits in the thread looking vindicated and inflates the heuristic that produced it. deployah on himself, 2026-08-24 SIGNET respin."
metadata: 
  node_type: memory
  type: feedback
  originSessionId: 125b9c1a-54cf-4f89-b6c2-bbc5e74892e8
  modified: 2026-08-25T01:04:22.321Z
---

**If the condition that would have tested your prediction never arose, the prediction was not
confirmed — it was never run. Say so explicitly, in the same thread where you registered it.**

deployah ruled this on himself, unprompted, 2026-08-24 (SIGNET #218 golden respin). He had
pre-registered that IF a second red landed at the fixed head, it would fall in the
never-yet-linted-on-Linux population rather than reading as "the fix is incomplete". No second red
arrived. His own words: *"A prediction whose trigger never fires is not a correct prediction, it is
an unexercised one, and I would rather log it that way than let it sit in the thread looking
vindicated."*

**Why it matters more than it looks.** Nobody goes back and audits a prediction that was never
contradicted. It stays in the record as a call that "held", and the NEXT time the same heuristic
speaks it carries borrowed credibility it never earned. This is the quiet inflation mechanism behind
[[identical-readings-across-opposite-outcomes-indict-the-meter]] and
[[dont-take-a-diagnosis-as-measured]]: a right-looking outcome that no instrument actually produced.
An unexercised prediction and a confirmed one are indistinguishable at the receiving end unless the
author labels them apart — and the author is the only one who cheaply can.

**What actually closed the question was measurement, not the inference.** The crate-level question
("is a second dead-code defect hiding behind the first?") could not be closed by the red run at all:
its lint pass aborted at the `spt` bin, leaving everything downstream unlinted. It was closed by
running both clippy arms at the fixed head and watching the ENTIRE workspace lint clean on both OSes
— 337 crates Windows / 326 Linux, zero warnings. **When a prediction and a measurement point the same
way, credit the measurement and say which one you are standing on.**

⚠ **The population you name in a prediction is itself a claim, and it can be wrong at the edges.**
doyle's fold-in widened deployah's residual: under `cargo clippy --workspace --all-targets`, a
dependency crate's **test and bench targets are independent leaf nodes**, so "that crate compiled
before `spt` failed" pins only its **lib** target — its test/bench targets may never have been linted
at all. So the never-linted population was larger than named. Before reasoning over "everything that
already passed", check what unit the pass was actually recorded at.

**How to apply:** when you pre-register a prediction, also register what would EXERCISE it. At
close-out, sort it into confirmed / refuted / **unexercised** by whether that trigger fired — never
leave it unsorted. Same discipline as [[declining-a-credit-specifies-the-requirement]]: naming the
condition your evidence never presented is what makes the evidence legible. Related:
[[a-frontier-names-where-progress-stopped-not-the-cause]].