Quoted Language Evidence
Tags: abbey-root • research • voice-analysis • evidence
Quoted Language Evidence
The first attempt to classify quoted language failed because 42 decisions were too much for the local model to return reliably in one response. The evidence session changed the unit of work instead of increasing the budget again.
The 165 candidates were divided into seventeen batches of no more than ten sources. Every response had to return every expected source exactly once with an allowed category and relevance code. A normalizer rejected missing IDs, extra IDs, invalid codes, malformed JSON, and duplicate coverage.
All seventeen batches passed without manual repair.
The complete distribution included 96 model-marked supporting cases, 65 comparisons, and 4 exclusions. Human review narrowed the retained core to 76 cases: 44 uses of scare or distancing quotation marks and 32 uses of invented labels or comic renaming.
Twenty direct-quotation or reported-speech cases remain provisional because their contribution depends on the surrounding authored frame. They were not silently merged into the core claim.
REVIEW-004 records all 165 decisions with the complete exact text from the frozen corpus. It passes deterministic review validation.
EVID-004 uses a much smaller canonical set spanning 2009 through 2021. It also retains ordinary song titles, direct attributed quotations, and reported speech as comparison evidence.
The resulting claim is narrower and more useful than “uses quotation marks.” The evidence supports quotation marks functioning as a stance marker: the writer signals that wording is disputed, nonliteral, inadequate, or being replaced with a comic alternative.
OBS-004 now has evidence, but no hypothesis or validation. The next step is to turn that narrow stance-marker claim into something testable.