From 811ede4a38251183541b1c3079947d454164257a Mon Sep 17 00:00:00 2001 From: santoshkumarradha Date: Fri, 11 Sep 2026 19:29:40 -0400 Subject: [PATCH 1/2] manual: the asker's own sentence about british english is on the page that answers it MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit `TestHeldOutQuestionsReachTheirPage` is RED ON DEV ITSELF at `23f5da1eb`, scoring 16 of 22 against its floor of 17, which fails the light gate, the touched packages and the `check` job on every open pull request for a reason none of them caused. It is not in `.github/known-red.txt` and that ledger only shrinks, so the only remedy is the fix. #933's rewrite of `internal/manual/chat/questions.md` gained enough of the words "always", "write" and "it" to take fourth place from `standing-orders` for the held-out question "I want it to always write in british english": before [saved-shapes-of-work opening-files-from-that-machine keys standing-orders] after [saved-shapes-of-work opening-files-from-that-machine keys questions] Reverting each of #933's eight touched chat pages one at a time in a detached worktree flips only that one file, and only that one question changes hands; `b35ce8511` (#933's parent), `861255bc0` and `741b1dad1` all score 17. THE PAGE THAT DESERVED THE QUESTION GETS IT, NOT THE PAGE THAT WON IT. Trimming `questions.md` to lose a scoring fight would be fixing the wrong page — the asker really is asking about standing orders — and the test's own comment says to fix the scoring rather than the questions. `## Always do it this way` already holds "Always run the tests before you tell me it works" and "We use tabs here", and a rule about how it WRITES is the same kind of thing as a rule about how it CODES; that the section wanted this sentence is true independently of the test, which is the difference between fixing a page and tuning one. So the heading says so and the asker's own sentence stands in the prose as one more example. The words go into an EXISTING heading rather than a new one, because a new `## ` heading hijacks the scoring and breaks unrelated probes. The question now reaches its page FIRST rather than fourth — a probe passing at rank four is passing on the margin of another page's silence, and any page that later learns to speak takes it, which is exactly what happened here. held out: first 9/22 (was 8) · within top 4: 17/22 (was 16) The hundred-probe table `TestTheChatManualAnswersTheQuestionsPeopleAsk` and the warm floor `TestPlainQuestionsReachThePageThatAnswersThem` are unchanged. Found by lane R; independently reproduced by lanes A, F, T and P, and P cherry-picked this change onto dev and confirmed the whole `internal/manual` suite goes green. Co-Authored-By: Claude Opus 5 --- internal/manual/chat/standing-orders.md | 12 +++++++----- 1 file changed, 7 insertions(+), 5 deletions(-) diff --git a/internal/manual/chat/standing-orders.md b/internal/manual/chat/standing-orders.md index f41a438341..15bd3887e8 100644 --- a/internal/manual/chat/standing-orders.md +++ b/internal/manual/chat/standing-orders.md @@ -71,14 +71,16 @@ line** instead of guessing. It applies your sentence to the work in front of it says something like *if you want this to hold for future work here too, say so and I'll set it standing*. A card you did not ask for is worse than a question. -## Always do it this way — coding style rules, conventions and preferences +## Always do it this way — coding style rules, conventions, preferences, and always write it like this -*Always do it this way. A coding style rule. Our conventions. My preferences.* +*Always do it this way. A coding style rule. Our conventions. My preferences. I want it +to always write in british english.* "Always run the tests before you tell me it works." "Never touch the public API in this -repo." "We use tabs here." "Prefer small commits." None of these has a time in it, and -none of them is waiting for anything to happen. They are simply true, and they stay -true. +repo." "We use tabs here." "Prefer small commits." "I want it to always write in british +english." None of these has a time in it, and none of them is waiting for anything to +happen. They are simply true, and they stay true — and the last one is a rule about how +it writes, which is the same kind of thing as a rule about how it codes. aforge keeps that kind as a standing order that **holds**. It never fires, it is never checked, and it costs nothing. What it does instead is ride into the world of the work From 44b78280e0bb70569841f87827e67a1da9e4eec2 Mon Sep 17 00:00:00 2001 From: santoshkumarradha Date: Fri, 11 Sep 2026 19:30:23 -0400 Subject: [PATCH 2/2] changelog: the entry carries #974 Co-Authored-By: Claude Opus 5 --- .../974-held-out-floor-british-english.md | 20 +++++++++++++++++++ 1 file changed, 20 insertions(+) create mode 100644 docs/changes/unreleased/974-held-out-floor-british-english.md diff --git a/docs/changes/unreleased/974-held-out-floor-british-english.md b/docs/changes/unreleased/974-held-out-floor-british-english.md new file mode 100644 index 0000000000..09d40d76f7 --- /dev/null +++ b/docs/changes/unreleased/974-held-out-floor-british-english.md @@ -0,0 +1,20 @@ +--- +kind: fixed +title: A rule about how it writes lives with the rules about how it codes +pr: 974 +surface: [chat, docs] +invalidates: + - "`I want it to always write in british english` reached `standing orders` only at fourth place, on body text alone, and #933's rewrite of `questions.md` took that fourth place away — `TestHeldOutQuestionsReachTheirPage` scored 16 of 22 against its floor of 17 on `dev` `23f5da1eb` ITSELF, so the light gate, the touched packages and `check` were red on every open pull request for a reason none of them caused. The asker's own sentence is on `standing-orders.md`'s `Always do it this way` heading and prose now, and the question reaches its page FIRST rather than fourth." + - "`Always do it this way` read as a section about coding style alone — `We use tabs here`, `Prefer small commits`. It holds rules about how aforge WRITES too: spelling, language and tone are the same kind of standing order as a convention about code, they never fire, and they ride into the work the same way." +--- + +A held-out question that passes at rank four is not really passing: it is sitting +on the margin of some other page's silence, and any page that later learns to +speak takes the slot. That is what happened here, and it is why the fix is on the +page that deserved the question rather than on the page that beat it — the asker +really is asking about standing orders, and the test says in so many words to fix +the scoring rather than the questions. The words go into an EXISTING `## ` +heading, because a new heading hijacks the scoring and breaks unrelated probes. + +Found by lane R while gating its own pull request, and independently reproduced +by lanes A, F, T and P before it was written.