The example sentences in VerbDetailView (and the new vocab practice
modes) frequently used the wrong verb — a "tener" set would show
"Él estaba leyendo un libro" (estar), "Nosotros vamos a viajar" (ir),
"Tú debes estudiar" (deber). The model drifted off the target verb
partway through generating the 6-example batch, and nothing checked
the output.
Two defenses:
Prompt grounding — VerbExampleGenerator.generate now takes a
formsByTense map (tenseId → conjugated forms, from the new
ReferenceStore.conjugatedForms). Each tense line in the prompt
lists the verb's exact conjugated forms and instructs the model to
use one of them. The model echoes a real form instead of recalling
(and mis-recalling) the conjugation.
Output validation — every generated sentence is checked against the
conjugation table via accent/case-folded whole-word matching. Any
sentence that doesn't contain a real conjugated form of the verb is
rejected. Failures trigger one regeneration pass; anything still
wrong is dropped rather than displayed. Better to show 4 correct
examples than 6 with 2 wrong.
Cache invalidation — VerbExampleCache now persists a versioned
wrapper (version 2). Pre-fix cached example sets — which may contain
wrong-verb sentences — fail the version check and are discarded, so
they regenerate cleanly under the new path.
Callers updated: VerbDetailView, VocabFlashcardPracticeView,
VocabMultipleChoicePracticeView all build formsByTense from
ReferenceStore and pass it through.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Trey T
f0eb75a28a