Bendik Aagaard LynghaugandClaude Fable 5.1
4d22f4cb44
Re-planning the contractor showed three wasted rounds: `moves_by: {
checking: submitter }` written as a bare word, a group without a
label, and the two featured owner roles left with nothing to play,
which the builder only warned about after the fact. The bare word and
the missing label now parse (a label defaults to the id), and a plan
that parses with spectator roles is sent back once with the finding;
on the contractor, round three fixed both.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2.4 KiB
The contractor, planned end to end, 2026-09-24
The first site made with no hand in it after the intake: plan with
Sonnet against the reference that now teaches the opener, the sender,
the mails and the guards; the ten held-out construction personas
mapped to the plan's need ids by a second model call (ids only, after
the plan was final); build with Sonnet writing the words and iris
green; then the visitor.
- Plan: 5 needs, 3 groups, 4 buckets, an opener for the two owner roles, a sender, mails on four needs. The builder warned twice, and rightly: neither featured owner got a move after sending. That is the plan model's omission and the reader's to fix before shipping.
- Score with descriptions, three seeds: 7, 7, 6 of 8. The hand-written contractor plan of 2026-09-21 scored 8 of 8 on the same personas, one seed. Inside the spread of each other; the planned site is as good as the hand-planned one on routing, and it took one command.
- Not done here: rating, and the headings-only pass, which the noise floor says would not have told us anything.
Rated
Thirteen mails, LIX 13 to 31, all very easy or easy, 23 to 39 words.
Every page: which, where, asked and next at 2 of 2 for every persona,
every control passing, prose at LIX 19 to 24. The budget audit found
four of five needs with max_fields equal to their required fields,
the tell of a model counting the form; only the crew's report (gloves
on, 3 tolerated, 2 asked) was decided from the intake. plan now
says so, and the reference tells the model to decide the number
first.
Re-planned twice, 2026-09-24
r2/: the reference now teaches the opener, the mails, stage words
and turned_down_from. The model used the opener, the sender and
mails again, ignored the words and the turned-down key (its ids are
English, so they cost nothing), lost round one to moves_by: { checking: submitter } written as a bare word, and again gave the two
featured owner roles no move. Desks: 19 of 19, English ids scoring
fine. r3/: the bare word is accepted, and the spectator finding goes
back to the model for one more round: round two flagged both roles,
round three fixed them, and the built site raised no warning. A group
without a label cost a round too; the label now defaults to the id.
The budget tell fired on both: three or four of the needs with
max_fields equal to the form. The model does not decide that number
first however it is told; the reader does.