Bendik Aagaard LynghaugandClaude Fable 5.1
8e6a93ebb7
The first desk score offered every button of the bucket, where portal shows a row only the moves from its own state; that scored the live vel 17 of 26 and sent a rewrite of the neighbour-help desk chasing an instrument fault. With the row's buttons only: 24 of 26, and the rewrite loses on the fair test too. Noted: an accent-stripped state id (ikke_lost) reads as English even to a strong picker. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
3.1 KiB
Tomter Vel: the doors stay, the opener comes first, 2026-09-23
The flat front door (runs/tomtervel-flat-2026-09-23) scored worse
than the live doored site: ten offers on one page put the close pairs
next to each other. So the plan keeps its four doors and gains what
the flat build was for: the opener (two cards, membership and
neighbour help, each with a move of the person's own afterwards), the
look-around (the three public lists), the mails on the stages that
matter to the sender, and the member's own way out. Words lifted from
the live repo by card identity - the lift now follows a card across
pages - with the opener in Norwegian by hand (opener-only.json).
Scored with the scorer that treats a card with no button as what the
page says, not as an option (iris after 2026-09-23). The first run
(r1/) lifted from the wrong checkout and had the scorer counting
opener picks as misroutes; it is kept for the record.
| Front door | With descriptions | Headings only |
|---|---|---|
Live, four doors (tomtervel-cards-2026-09-23) |
32 of 40 | 29 of 40 |
Flat, ten offers (tomtervel-flat-2026-09-23) |
12 of 20 | 9 of 20 |
| Four doors with the opener, this run | 30 of 38 | 26 of 38 |
Level with the live site within the visitor's noise, with everything new on top. Merged to tomtervel/questions main the same evening. The mails are the builder's fallbacks; a model has not reworded them.
Round 2: the mails, in Norwegian
build --llm-for aggregates.yaml with Sonnet, every other word as
merged. Six mails, all in Norwegian, both links to the sender's own
move kept (the builder puts a lost link back). Readability level with
the fallbacks, LIX 7 to 33 against 13 to 20, and the text says what
happened: "Vi har funnet en nabo som kan hjelpe. De tar kontakt med
deg direkte." On vel.klingenbergbygg.no the same evening. The builder
now refuses a reply in the wrong language for the plan and keeps the
words it had.
Round 3: the doors reworded, and rejected
Sonnet rewrote the four door cards (words-r3.json, site-r3/),
and the result was scored with descriptions across three seeds
against the live doors: 29, 25, 28 of 38 against 30, 29, 29. Worse,
outside the spread. "Få hjelp av naboer eller gi en hånd selv" takes
the volunteers as well as the neighbours, because "gi en hånd selv"
reads as volunteering. The live doors stay. What survives every round
is the same four: hearing input read as a report, a committee read as
a hearing, and volunteering read as neighbour help - which says the
doors' seams are the plan's, not the words'.
Round 4: the neighbour-help desk reworded, and rejected
The first desk score said this desk got 2 of 4 buttons right. Sonnet
reworded its labels (/tmp only); the score got worse with both a
small and a strong picker, and the strong one showed why: told a case
"has just become ikke lost", it pressed "Meld ordnet" - it read the
accent-stripped state id as English. The scorer also offered every
button of the bucket where a row shows only its own state's. With the
row's buttons only, the live desk scores 3 of 4 and the vel 24 of 26
over all desks. The labels stay as they were.