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
39 changes: 39 additions & 0 deletions .claude/skills/ci-gates/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -941,6 +941,45 @@ be invented. The temptation to produce a diff is strongest exactly when the
honest result is "the plan was wrong" — and a diff manufactured then is the
purest form of the thing this whole skill is about.

## 24. Check the regression before proposing the next thing

Seventy merges landed in the twenty-four hours after this campaign's work. None
of them was reviewed against the numbers the campaign produced. **Re-running the
measurements is the first item of the next iteration, not an afterthought** —
and it is the only way "the new cycle does not break the old work" is a fact
rather than a hope.

Everything held. The detail worth keeping is *why* the most fragile piece held:
`check_gate_preconditions.py` names its two uncovered branches **by message**,
and other people's edits moved those lines from `:359/:403` to `:451/:495`
without breaking the declaration. A line number in a note is a claim with an
expiry date nobody sets.

### The decision brief is part of the work, not a wrapper on it

Eight questions accumulated that are an owner's to answer: repository security
settings, deleting reproducibility records, corpus-wide language decisions. Left
as eight issues they are eight interruptions; collected with **options and the
cost of doing nothing**, they are one sitting.

Two properties make such a brief usable rather than decorative:

- **Every number comes from running a command**, named in the footer so it can
be re-run. A brief that summarises other summaries inherits every error in
them and adds none of the evidence.
- **It says how often its author has been wrong.** Seven claims in this campaign
were published and then refuted by my own later measurement. A decision
document whose author has that record should carry it, or the reader has to
discover the reliability of the source by accident.

### When "fix it" is not on the menu, say what is

`coverage` is red for a real reason and cannot be moved by bookkeeping *or* by
one compiler fix — re-sealing moved it zero, and the backlog is zero specs deep
at depth 1. The honest options are a multi-month body of work or a decision to
call it accepted debt. Offering a third, cheaper-sounding option would be the
comfortable thing and there isn't one.

## 25. A note naming what you did not cover is a work item

`check_gate_preconditions.py` said, of two branches it left uncovered:
Expand Down
15 changes: 15 additions & 0 deletions docs/now/2026-08-24-eight-decisions-that-are-not-mine.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,15 @@
# NOW — eight decisions that are not mine (2026-08-24)

The gate audit has produced eight questions I deliberately did not answer: repository security settings, deleting reproducibility records, and corpus-wide decisions about the language. Each now has options and a stated cost of doing nothing, in one place.

- **Regression checked first.** After **70 merges in 24 hours**, everything built in this campaign still holds on fresh master: 12 of 13 gates kill every mutant, the two survivors are exactly the pair declared `UNCOVERED`, and `verify_exhaustive` (the gate that was permanently red until yesterday) exits 0 with its control green. The `UNCOVERED` note still matched after other people's edits moved those lines, because it names branches **by message rather than by number**.

- **No duplication this time.** Before proposing a "regression command", I checked `t27c gates` and `t27c battery`. `gates` runs every `check_*.py` and reports each child's own exit code, and separates scripts that cannot fail; `battery` is for document reconstruction oracles. Neither runs the negative controls, which is what `tri gates sweep` and `tri gates mutate` add. Complementary, not overlapping — the first time in four iterations that the §21 check came back clean.

- **The brief.** Eight decisions, grouped: repository settings (#2455, #2474), reproducibility records (#2477), language and corpus (#2479 and its link to the parse backlog), and four small measured defects (#2476, #2483, #2466, #2472). Two of them — #2476 and #2483 — need the stage0 freeze lifted, which is why they are filed rather than fixed.

- **The one connection worth reading twice.** `coverage` cannot be moved by bookkeeping *or* by one compiler fix. Re-sealing the 18 stale twins moved the count by zero because 14 of them do not parse, and `t27c backlog` reports zero specs at depth 1 with 94% four or more defect classes deep. It is a multi-month body of work or a decision to call it accepted debt — and those are the only two honest options.

- **Seven claims published and then refuted by my own measurement** over this campaign, every correction appended to the issue that carried the claim. The brief says so in its footer, because a decision document whose author has been wrong seven times should say how often and where.

Refs #2455, #2466, #2472, #2474, #2476, #2477, #2479, #2483
Loading