Every prompt and reply is now kept under <out>/llm/. An empty reply is
a retry, not a fatal error; a spec that never parses stops the run
before the structure phase; --visitor-llm lets a different model play
the visitor. The file-block parser finds markers anywhere in a line,
accepts a short END and lets a new FILE close an open one, after a
reply glued two files together with `<<<END>></think>`. The reference
gains a language section: copy in the intake's language, identifiers
in ASCII.
runs/rehab-entreprenor-no-2026-09-21: nemotron-3.5-lightning:30b-a3b
did not produce a loadable site. Written up with the reason and the
way forward: structure from needs.yaml deterministically, words from
the model.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
Readability is counted (LIX), not judged. Understanding is three
yes-or-no questions per persona, with a gutted twin of every page as
the control: an instrument that still says yes to "Option 1, Option 2"
is not measuring the page. Friendliness is a preference between two
versions shown in both orders; a pick that flips with the order counts
as no preference. Page load (options offered, required fields in the
step) flags a page that should be split into a further question or
have fields moved to a later stage.
First run on the generated contractor site, qwen3:8b as the visitor:
control passed on every need; "what happens next" was 0/2 everywhere
and 2/2 everywhere once the page said who reads it and when (LIX 28 to
30); an official-voice control reads LIX 63 against 27. Friendliness
mostly came back "no preference" at two personas per need: directionally
right (never preferred the cold version), too thin to rely on.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
`portal-trainer simulate` reads the model `question_lint --needs-sim`
exports. A visitor model picks an alternative and fills the form in a
manner that varies by run; the case is then walked along the handling
group's own desk buttons to a finished state, a different one each
run, so every state and button is exercised. At each button the
handler model is asked one thing: with only what the form collected,
could this role really do what the button claims? Its question goes
back to the visitor, who may not know.
Who wins is deliberately not simulated. The first version let the
handler decide: qwen3:8b sent every case to "won", and after one
sentence of prompt change, every case to "lost". Outcomes belong to
the live bucket; `success` in needs.yaml is what a report over real
records counts.
`compare` sets two score files side by side. engines/ollama.sh honours
SEED, FORMAT=json and TEMPERATURE so the same case replays the same
way against two versions of a site. generator/reference.md documents
`success` and `stages`.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
runs/contractor-2026-09-21: the generator's first end-to-end run, on
intakes/contractor.yaml. Needs parsed first round, structure green
first round, 10/10 on its own personas - and 8/10 headings-only on
ten held-out personas with qwen3:8b as the visitor, because the pages
had been written with the personas in view. Phase 2 now shows the
model the needs without the personas.
`portal-trainer answer` exposes the blind answerer on its own. README
rewritten around what this is: an intake in, a verified content repo
out.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
`portal-trainer generate` turns a plain-words business intake into a
content repo in three phases, each held to portal's question_lint:
needs.yaml (then frozen, so the generator cannot weaken the spec to
pass), structure (revised against the lint's findings until green),
and wording (personas answered blind, misroutes fed back; a rewrite
that breaks structure is reverted). Model output is written to disk
only under aggregates.yaml, site.yaml and questions/**.yaml.
`portal-trainer answer` is the blind answerer alone.
The LLM is an external command, prompt on stdin, text on stdout:
engines/ollama.sh (local Ollama), engines/anthropic.sh (API key),
engines/mailbox.sh (hand each prompt to an agent or a person).
portal-portal/ is the onboarding site as a portal content repo: the
front door is the intake interview, records land in `onboardings`
(open -> drafted <-> revising -> live | declined), and it passes its
own needs spec.
runs/uhhm-2026-09-21/ollama: qwen3:8b and granite4.1:8b as blind
answerers on the same eight tasks.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
The trainer procedure as an agent skill, the engine contract, two
sector specs (software-studio: UHHM's own, green; construction: a
target with no pages yet), and the first wording run against uhhm.no.
The verifier itself is portal's question_lint (uhhm/portal 188f9bf).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L4jrCgLiKKHAEFuZUJjckH
