The Verb That Vanished
Can a translation drop a distinction that its own commentary names—and can revision restore it without adding something new?
This is a record of one selected conversational anomaly and a prepared pilot, not a report of completed model experiments.
Three outputs. Not three proofs.
Visible order within one response. Short excerpts manually transcribed from the conversation; not a provider raw-log export.
Source phrase: piétiner un peu.
“keeping the imagery and the cruelty of the original intact”
“even if it means stepping, just a little, on the one who's been there for so long.”
“As though trampling the original occupant of the cross is a minor inconvenience in the pursuit of visibility.”
Different word choices in translation and commentary, alongside an assurance about fidelity.
A candidate loss of force and a candidate overclaim. Literary adequacy remains a judgment, not a keyword match.
Prior understanding, intent, recurrence, internal cause, or superiority of older translations.
Later recognition does not prove earlier understanding.
The discovery is post hoc and context-rich. Further replies in this thread are exposed to the hypothesis; they are not independent replications.
Preserve the source. Do not maximize the violence.
The dictionary distinguishes physical piétiner from stalled progress. That makes “always say trample” a bad repair. A plain source can need plain English; a negated action must stay negated. Sources S1–S2.
Translate → assess
Establish a translation without a preceding annotation task.
Translate → translate again → assess
Control for an ordinary second attempt.
Translate → annotate → revise → assess
Separate recognizing a source feature from retaining it.
Annotate → translate → assess
Give feature identification an opportunity to precede output.
Planned: 48 independent sessions, 144 public response turns, one replicate per item/condition. None have run. Extra annotation also adds a turn and tokens; this pilot cannot isolate internal mechanism or analysis content from additional generation budget.
Source cards & clean initial requests
Operator view. These are original synthetic sentences, not Camus quotations. French naturalness and the feature notes still need independent bilingual review. Do not send this page or its notes to a responder.
Show author’s feature notes (not participant material)
No provider is connected. This page does not collect model responses.
No preferred-word scoreboard.
Review source and anonymized translation before seeing condition, annotations or self-assessments. Keep propositions, force, register, literary contrast, and negation/attribution as separate judgments.
Then compare source-feature recognition and claims of fidelity with those judgments. Preserve uncertainty and reviewer disagreement.
Two independent bilingual reviewers are planned; zero ratings exist. Elicited self-assessments are not a measurement of spontaneous assurances.
What is ready
The discovery record, twelve source cards, four-condition protocol, randomized session plan, and standard-library prompt/review preparation tool.
What is not established
Fresh model execution, source/rubric validation, translation ratings, adoption, or Windows-folder persistence.
17 / 17 software checks passed.
Fixture-based checks only. Not 17 model trials.
Evidence, not an authority costume.
S1. Académie française — piétiner, ninth edition. Physical/transitive uses and intransitive non-progress. Accessed 25 October 2026.
S2. Larousse — piétiner. Physical action and stalled progress. Accessed 25 October 2026.
The dictionaries do not validate literary quality or model-behavior hypotheses. Edition-specific translation comparisons from the conversation are excluded from this pilot’s evidence.