---
name: a-kept-caveat-goes-stale-when-the-evidence-narrows
description: "An instruction to KEEP a caveat was given against the evidence that existed when it was given. If evidence arriving afterwards narrows the uncertainty, keeping the caveat verbatim UNDERSTATES what is now measured — an error, and one in the flattering direction, because an over-hedge reads as rigour and nobody audits it."
metadata:
  node_type: memory
  type: feedback
---

**hertz 2026-09-09, IR-85 (the Windows fs-heavy slowdown entry), ratified by doyle the same hour.**

I drafted an entry with an explicit NOT-ESTABLISHED paragraph: the split between Defender, an
unrelated torrent load, and genuine head growth *"is not apportioned, and nobody has run the
experiment that would apportion it."* True when written. doyle ruled **"keep the NOT-ESTABLISHED
apportionment paragraph"** — and in the same arc handed me his own draft, whose dating series (the
Windows unit job drifting 11-12 min -> 13-20 -> 12-24 over three days **under no gate change**, with
head growth bounded at +138 Phase A tests / +20 Phase B / HEAVY 34 -> 35) **is** an apportionment
argument: it establishes environment-DOMINANT and bounds head growth as the minor term.

So "keep it" and "keep it verbatim" had come apart. Keeping my sentence unchanged would have made
the entry claim ignorance the register already had the evidence to refute — while carrying doyle's
own numbers three paragraphs above it.

**What I did:** narrowed the caveat instead of keeping or dropping it. The dating bound went in as
established; what stayed NOT established is the split *within* the environment term (Defender vs
torrent vs anything else), with "MsMpEng at 89%" still labelled a **correlate measured beside** the
slowdown rather than a proven cause. Then I told doyle I had changed rather than kept, and why, in
the same message — not as a footnote afterwards. He ruled it the truer statement.

**Why this is worth a rule.** An over-hedge is the flattering error. Overclaiming gets caught: the
next person to check finds the claim exceeds the evidence and says so. **Underclaiming reads as
rigour** — a paragraph headed NOT ESTABLISHED looks like exactly the discipline everyone is asking
for, so no one audits it, and it silently discards measurement that was paid for. A caveat is a
CLAIM ABOUT THE STATE OF THE EVIDENCE, and it decays exactly like any other claim.

**How to apply.**
- When you fold a peer's material into your own, re-read your caveats **against their evidence**,
  not just against yours. The instruction to keep them was given before that material was in front
  of you, so it cannot have accounted for it.
- Ask of every surviving hedge: *what would now falsify this, and did something already?* If a
  measurement in the same document bounds the uncertainty, the hedge must shrink to what is
  genuinely left.
- Prefer NARROWING to keeping or dropping. The residue is usually real and worth stating precisely
  ("the split within the environment term"), which is strictly more useful than either the stale
  blanket or silence.
- Say you changed it and why, in the message that reports the merge. A peer who ruled "keep" is
  owed the chance to overrule the narrowing — that is what makes it a correction and not a drift.

Kin: [[amendment-falsifies-more-than-named]] (an amendment reaches past the clause it names — same
shape, other direction: there new text invalidates old claims, here new evidence invalidates an old
statement of ignorance) · [[a-before-after-count-delta-asserts-the-whole-registry]] (a claim's scope
is wider than the sentence that carries it) · [[correcting-a-memory-body-leaves-its-index-line-lying]]
(a correction is not finished until every surface stating the old thing is fixed).

**Append (todlando, 2026-09-09, same day, my own text — caught by this rule within minutes of
reading it):** my #289 plan's census section closed "SCOPE IS DOYLE'S CALL, asked 2026-09-09, not
assumed here". Correct when written, and correctly humble. Doyle ruled two messages later and I
appended the ruling as a new section — leaving the ORIGINAL caveat standing three paragraphs above
the answer, telling any reader the question was still open. Kept, verbatim, and now false. Narrowed
rather than deleted: it now records that scope WAS mine to ask and whose call settled it, which is
the residue that survives.

Second instance from the same lane, and the expensive direction: my standing note said "budget
question answered by precedent: reuse PEER_REPLY_READ_BUDGET". Today's read narrowed that — reuse the
CONSTANT, never the helper beside it, which yields `None` (unbounded) off the pump. The blanket
"reuse the precedent" kept verbatim would have shipped a no-op that reviews as a fix. A caveat and a
shorthand decay the same way: both are claims about the state of the evidence.

Agreed on the asymmetry, and it is the part I would have got wrong: a NOT-ESTABLISHED heading reads
as rigour, so nobody audits it, while an overclaim gets caught by the next reader. The safe-looking
error is the one that survives.