Discovery audited · Protocol not run · Exhibit added 2026-10-26
The Verb That Vanished
A translation and later commentary disagree about a source-language distinction. What experiment could establish whether that matters?
Open the interactive exhibit ↗The selected observation
One post-hoc, manually transcribed conversational excerpt shows different wording in an initial translation and subsequent commentary, alongside an assurance of fidelity. This is neither a provider raw-log export nor a controlled replication. The interesting question is whether preserving source meaning differs from noticing it after the fact.
The planned comparison
Visitors can inspect twelve author-created French passages and four planned trial conditions: direct translation, simple retry, analysis before, and analysis after. The page generates initial prompts and exports the planned protocol; it does not run a model or collect replies. The packet specifies 48 isolated trial sessions and independent bilingual review, but there are zero fresh model trials and zero independent ratings.
What would count
Bilingual reviewers would need to assess propositions, register, force, attribution, negation, and amplification without seeing the trial condition. Literal word matching alone cannot establish fidelity. A second analysis step also costs extra tokens and generation turns, so the design does not cleanly isolate an internal mechanism.
Checks and unknowns
Seventeen local Python fixture tests re-ran and passed on October 26; they test protocol setup and data isolation, not translation quality. An injected-document Chromium smoke test exercised prompt selection and JSON export without page errors. Ordinary hosted navigation, native Windows clipboard and independent review were not established.
Artifact receipt
Archived original standalone HTML, preserved without source modification. SHA-256: 5159ceb535967237b0fe8af5b9f7d660ad4b43636a9fad07ca58f01c1b24a18a. Byte identity is not independent scientific validation.
Gallery navigation update · 2026-10-27: The public instrument adds only a scoped navigation bar and its styles. Scientific scripts, embedded data, and original instrument content are unchanged. The historical hashes identify the archived standalone original, not the navigation-adapted index.html.