Skip to content

docs(skill): audit the skill itself; one retraction was live, one number stale (#2492) - #2497

Merged
gHashTag merged 2 commits into
masterfrom
docs/audit-my-own-skill
Aug 23, 2026
Merged

gHashTag merged 2 commits into
masterfrom
docs/audit-my-own-skill

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

Yesterday's finding was a justification I had written a day earlier that sounded principled and was wrong. Thirty-three sections were written the same way, in the moment, and none had been re-read with that question.

What was checkable, and what it said

result
ten file references all exist
four behavioural assertions all hold
five numeric pairs one stale
contradictions between sections one live

The contradiction is the finding

§32 stated, in the present tense, the justification that §33 retracts — and §32 comes first. A reader who stops there gets the wrong rule with a confident explanation attached.

It now carries RETRACTED, see §33 at the top of the paragraph, not a note appended after it. Appending is always the cheaper edit, and the reader who most needs the correction is the one who does not reach it.

The stale number

"34 of 42 sites" was true when written and false the next day: §33 tightened the predicate and the denominator fell to 36, because the loose one had been counting helper functions.

Dated in place rather than corrected. The sentence is about what the scanner missed at that moment; a number silently updated each time teaches nothing about how it drifted.

A near-miss worth recording

"30 of 109 catalog rows share a citation" read false at first — I counted 45. Both readings had to be computed before saying anything:

rows in shared groups           45
duplicates beyond the first     30   <- what a uniqueness check flags

The claim holds. Second time in three days that computing both readings prevented a false accusation against working code.

The general form, now §34

Every rule this skill has about gate output applies to the skill itself: dated measurements as standing facts, a label trusted over a property, a justification that sounds principled. The difference is that nothing runs prose, so nothing goes red when it rots.

Refs #2492

…ber stale (#2492)

Thirty-three sections written over six days, none re-read with the
question that catches a stale claim. Ran the audit on the file: ten file
references, four behavioural claims, five numeric pairs.

THE CONTRADICTION. Section 32 stated in the present tense the
justification that section 33 retracts, and 32 comes first -- a reader
stopping there gets the wrong rule with a confident explanation attached.
It now carries RETRACTED at the top of the paragraph rather than a note
appended after it.

THE STALE NUMBER. "34 of 42 sites" was true when written and false the
next day, because 33 tightened the predicate and the denominator fell to
36. Dated in place rather than updated: the sentence is about what the
scanner missed at that moment, and a number that keeps being silently
corrected teaches nothing about how it drifted.

A NEAR-MISS. "30 of 109 rows share a citation" read false at first -- I
counted 45. Both readings had to be computed: 45 rows sit in shared
groups, 30 are duplicates beyond the first, which is what a uniqueness
check would flag. The claim holds. Second time in three days that
checking both readings prevented a false accusation.

Section 34 records the general form: nothing runs prose, so nothing goes
red when it rots.

Refs #2492
@github-actions

Copy link
Copy Markdown
Contributor

📓 NotebookLM Notebook linked to this PR

This notebook contains session context, decisions, and artifacts for this work.

@github-actions

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-08-23 03:57:40 UTC

Summary

Status Count
Total Open PRs 6
PRs with Failing Checks 2
PRs with All Checks Green 4
READY 3
FAILING 2
PENDING 0

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=e905ba8eeb53 != manifest seal=87e5cbd3ad94.
    The committed NMSE numbers were certified against an older compiler.rs.
    Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).

@gHashTag
gHashTag merged commit 1321e48 into master Aug 23, 2026
18 of 21 checks passed
@gHashTag
gHashTag deleted the docs/audit-my-own-skill branch August 23, 2026 04:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant