Seen live: "Esta manhã acordei cedo e fui correr ao longo da marginal.
O ar estava fresco e havia poucas pessoas na rua." — unremarkable
Portuguese, two marker hits, zero English hits, and a verdict of
English. Corrected as English, read aloud in an American voice.
The list was missing the ordinary machinery of the language: the
contractions (ao, à, num), the tenses a diary is written in (estava,
havia, fomos), the words that join two clauses (até, depois, então,
onde). Each clears the bar the list already set — an English sentence
has no reason to contain them — so their absence bought nothing.
The floor stays at three. What changed is that three is now reachable
by prose rather than only by a paragraph that argues its own case. fr
and es get the same additions by analogy; neither has an account yet to
catch it live, which is exactly how this one survived. "sin" and "tan"
stay out of the es list: both are English words.
Two regression tests, pointed in opposite directions — ordinary
Portuguese must read as hers, and English about Portugal, English
quoting Portuguese, and a plain English diary must all still read as
English, held verdict included.
Claude-Session: https://claude.ai/code/session_01GJHNvirh7Hzhc9RL3HAvz7
Storing doc_lang on the document row is not enough on its own. The
editor sees that row when the document is opened or saved, and the pass
that decides the verdict runs after a save — so the client was always
one save behind, and read-aloud is reached for precisely when she has
stopped typing and no further save is coming. Heard in a browser: a
Portuguese paragraph read in an American voice, twice, until another
keystroke went in.
/check, /voice and /collocation now answer with X-Petal-Doc-Lang. A
header rather than a wider body: all three answer with a bare array of
the unified pending set and every caller reads it as one, and a verdict
is metadata about the pass rather than another suggestion. It reaches
the app through the same handler shape onUnauthorized already uses.
Claude-Session: https://claude.ai/code/session_01GJHNvirh7Hzhc9RL3HAvz7
Phase 28's step (b): both remaining items are about direction, and both had
a wrong answer that looked right.
isTranslation could not simply be read backwards. readsAsEnglish is a
deliberately low bar — Latin letters, not swamped by another script — which
every Portuguese sentence clears as easily as English does, so swapping its
two halves would have called every genuine Portuguese correction inside a
Portuguese document a translation. The flipped direction uses sentenceLang
from doclang.go instead, where English has its own curated marker list and
has to out-evidence the pair language to win. The English-document path is
untouched; reconcilePending carries the verdict to ask the question the
right way round.
The tap-through's whole observable change is a model call that stops
happening. /suggestions/{id}/translate now recovers the explanation's
language by re-running targetFor rather than assuming the pair, which gives
today's answer everywhere except the case that was broken: the Portuguese
writer whose explanation already arrived in Portuguese, previously
round-tripped through the model into Portuguese again. It answers "" there,
and the client's existing `res.translation.trim() || explanation` fallback
seeds the bubble with the explanation itself — no frontend change at all.
It deliberately does not render that explanation into English on the
grounds that English is technically the other half: an unasked-for
rendering into the language she is practising is noise, not a seed.
Tests pin both directions of the detector, with Portuguese-in-Portuguese as
the case the file exists for, plus four handler tests through the real
/check and /translate paths — including the skipped seed asserting the
model was never called, and the learning_pair zh learner whose English
explanation still renders into Chinese.
Left of the phase: (c) the garden's language tagging and read-aloud.
Claude-Session: https://claude.ai/code/session_01GJHNvirh7Hzhc9RL3HAvz7
Every pass was English-shaped: CheckpointMessages took the text and the tone and
nothing else, so there was never a language decision to get wrong. On the live
build two pt-PT sentences drew no cards at all — Petal read the Portuguese, said
nothing about it, and filed a mechanics note about the one English line.
The rule is two decisions reading different state. What gets corrected follows
the document. What language the explanation is written in follows the writer —
the half of her pair she is not learning, from users.direction — because an
explanation is teaching, and teaching lands in the language she reads most
easily. Those coincide for every account that exists today (learnerPairs is
{"zh"}), which is a fact about the roster and not about the design, so Target
keeps them apart. It carries a third language too: the collocation gloss is
addressed to her rather than to the document, and folding it into Explain would
have quietly moved it into English on every English document.
The document verdict is a proportion, not a presence — one Portuguese quotation
must not flip an English essay. Per sentence, three-way: pair, English, or no
answer. The third value is the load-bearing one; counting the undecided as
English is exactly what would hold a journal of short Portuguese sentences in
English forever, so the Latin pairs needed an englishMarkers list curated against
pt/fr/es as carefully as latinMarkers was curated against English. Hysteresis at
70/40 because a bilingual paragraph would otherwise alternate its cards' language
every few keystrokes, and hysteresis needs a yesterday — hence the column. Plus a
corroboration floor: a ratio computed over "Não. Eu." is 100% of nothing, and a
flip rewrites every card in the document.
The verdict folds into the chunk salt beside the tone, so a document that changes
language re-opens every sentence rather than serving back cards in a language it
no longer speaks.
checkpointSystemPrompt could not simply take a language — it opens by naming the
reader an ESL learner, and appending "explain in Portuguese" hands the model two
contradictory framings. Separate constants, sharing the JSON contract below the
framing. Both carry a "never translate it into English" line, which is the
instruction the model will most want to disobey. The English prompts are
untouched byte for byte, and a golden says so out loud.
Collocation deliberately did not move: its prompt is per-language knowledge, not
framing, and "natives usually say" for Portuguese is a claim Petal cannot back.
Not deployed and not smoked against a real model. The tests drive the real router
and a real DB; what none of them prove is how Qwen behaves on a Portuguese
document, in particular whether the never-translate line holds.
Claude-Session: https://claude.ai/code/session_01GJHNvirh7Hzhc9RL3HAvz7
prosolis