Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
69 changes: 60 additions & 9 deletions .claude/skills/ci-gates/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -662,8 +662,11 @@ a thing whose coverage nobody measures.
expressed, so those branches reported as uncovered while the file that
covers them sat in the tree.
3. **One syntactic form.** It matched a bare `return 1..4` and missed seven
ternaries and a `raise SystemExit(3)` -- 34 of 42 sites. It reported a gate
whose every verdict is a ternary as having "no failure path to break".
ternaries and a `raise SystemExit(3)`: 34 of the 42 sites it could see AT THE
TIME. It reported a gate whose every verdict is a ternary as having "no
failure path to break". That denominator is a **dated measurement** — §33
tightened the predicate and 42 became 36, because the loose one had been
counting helper functions.
4. **No baseline.** A mutant is killed when the control exits non-zero, so a
control that is red BEFORE any mutation scored a perfect kill on everything.
The exact inverse of the defect the tool exists to find.
Expand Down Expand Up @@ -1146,7 +1149,8 @@ cases demanding RED satisfied all of them, and the **opposite** defect went
unmeasured across thirteen gates.

`--loud` flips `return 0` to `return 1`: does anything notice a gate that fails
on a **clean** tree? First run, **seven of thirteen** have survivors.
on a **clean** tree? On its first run seven of thirteen had survivors; all are
closed now, which is a fact about a date and not about the tool.

The worst was the file that enforces this discipline for six other gates:

Expand Down Expand Up @@ -1241,13 +1245,16 @@ failure breaks any script reading it, and nothing would have said so.

### "No path to break" is unmeasured, not clean

Two gates report `no success path to break` under the loud operator: they hold
no bare `return 0`, every verdict being a ternary or a `SystemExit`. The
operator takes only bare returns on purpose, because forcing a ternary's whole
line to 1 is the silent operator seen backwards and would be scored against the
wrong control.
**RETRACTED, see §33.** This section originally said two gates were in that
state because "the operator takes only bare returns on purpose, since forcing a
ternary's whole line to 1 is the silent operator seen backwards". **That reason
was wrong and I wrote it.** Silent forces the line to 0 — the gate never fails.
Loud forces it to 1 — the gate always fails. With ternaries allowed one of the
two measures fine; the other has no literal success return at all, its success
being `return code`, so the row is true there for a different reason.

So those two are **unmeasured in that direction**, and the report has to say
The rule the section is for still stands: a row saying `no path to break` is an
**absence of measurement**, and the report has to say
which. A row that reads as a clean score when it is an absence of measurement is
the same substitution this whole skill is about — and it is easiest to commit in
your own tool's output, where you know what the row means and the next reader
Expand Down Expand Up @@ -1304,3 +1311,47 @@ before deciding whether to be pleased.
Two produced missing coverage. This one produced a row that **read as a score**,
which is worse: the first kind gets found by the next person who looks, and the
second gets quoted.

## 34. Audit your own skill the way you audit a gate

Thirty-three sections were written over six days, each in the moment, none
re-read with the question that catches a stale claim. So I ran the audit on the
file itself: every named file, every behavioural assertion, every number that can
be checked today.

**Ten file references, four behavioural claims, five numeric pairs. One was
stale, and one contradiction was live.**

- **The contradiction is the finding.** §32 stated, in the present tense, the
justification that §33 retracts — and §32 comes first. A reader stopping 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.

- **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 point of the sentence is what the
scanner missed at that moment, and a number that keeps being silently corrected
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: 45 rows sit in shared groups, and **30 are duplicates beyond the
first**, which is exactly what a uniqueness check would flag. The claim holds.
Checking both readings is the difference between a correction and a false
accusation, and this is the second time in three days that habit saved one.

### Prose you wrote is a claim like any other

Every rule this skill has about gate output applies to the skill: dated
measurements presented 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.** The only mechanism is re-reading it
against the tree, on purpose, with the same presumption of being wrong.

### A contradiction resolved in place, never appended

Repository doctrine here is to fix a contradiction where it lives rather than
adding a second truth next to the first. That is easy to honour in code and easy
to forget in a document, where appending is always the cheaper edit — and where
the reader who most needs the correction is the one who stops before reaching it.
17 changes: 17 additions & 0 deletions docs/now/2026-08-24-auditing-my-own-skill.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,17 @@
# NOW — auditing my own skill (2026-08-24)

Yesterday's finding was a justification I had written a day earlier that sounded principled and was wrong. Thirty-three sections of the `ci-gates` skill were written the same way, in the moment, and none had been re-read with that question. So the audit ran on the file itself.

- **What was checkable.** Ten file references, four behavioural assertions, five numeric pairs. Everything else is judgement, and judgement is not what rots.

- **All ten files exist. All four behavioural claims hold** — `check_duplicate_agreement` really does declare two control flags; the three import relationships are real.

- **One contradiction was live, and it 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 rather than a note appended after it, because appending is the cheaper edit and the reader who most needs the correction is the one who does not reach it.

- **One number was stale.** *"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, and a number silently updated each time teaches nothing about how it drifted.

- **A near-miss.** *"30 of 109 catalog rows share a citation"* read false at first: I counted 45. Both readings had to be computed before saying anything — 45 rows sit in shared groups, and **30 are duplicates beyond the first**, which is exactly what a uniqueness check would flag. The claim holds. Second time in three days that computing both readings prevented a false accusation of working code.

- **The general form.** Every rule this skill has about gate output applies to the skill: 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
Loading