Quoted Language Validation
Tags: abbey-root • research • voice-analysis • validation
Quoted Language Validation
VAL-004 tested the stance-marker hypothesis without reusing the canonical EVID-004 examples.
The deterministic definition removed fifteen supporting examples, seven comparison examples, and four rejected matches from the 165 reviewed quote-bearing candidates. The resulting holdout contained exactly 139 sources.
Sixty-one holdout posts used the two core functions, a rate of 43.88%. The 95% Wilson interval was 35.91% to 52.19%, leaving the lower bound above the 30% threshold.
Both subtypes replicated independently: thirty-six distancing cases and twenty-five comic-renaming cases remained after canonical evidence was removed.
The pattern also survived the chronology test. Core cases appeared in all four required bands, although the count declined from twenty-five in 2009–2011 to only three in 2018–2021. There were no holdout core cases after 2021.
Sixty ordinary quotation comparisons remained. That boundary is important because it demonstrates that the method is not simply labeling every quotation mark as voice evidence.
Every threshold and source-integrity check passed. HYP-004 is provisionally supported with medium confidence inside the tested Facebook scope.
The fourth chain is now complete. The next useful step is not automatically creating OBS-005. It is synthesizing the four validated chains into a bounded first Voice Model draft and being explicit about what the Facebook research does not yet establish.