Put three verdicts on “professional” and “complete” side by side. Local NLP extracts outcomes, obligations, negations, and suspiciously shared language before you accept a majority.
Waiting for three rulings.
No sentences analyzed
No sentences analyzed
No sentences analyzed
Computed panel result
No panel analysis yet.
Load the worked dispute or add three substantive rulings.
Agreement-Outcome tally
Highest overlap-Pairwise Jaccard
Correlation caution-Shared language check
Obligation conflicts
Extracted must, shall, required, may, and no-obligation sentences appear here.
Pairwise language
Normalized overlap and longest shared phrases appear here.
Boundary: this is language analysis, not a truth oracle or a reproduction of GenLayer consensus. Similar wording can indicate shared evidence, boilerplate, or correlated reasoning; inspect provenance before treating votes as independent.
Why this matters
Independence is a property of the reasoning path, not the vote count.
Three verdicts can still behave like one opinion when they inherit the same prompt, source, model family, or copied rationale.
The workbench separates outcome agreement from language overlap. That distinction does not prove independence, but it reveals where a validator should ask for provenance, diversity, or a genuinely fresh appeal.
An appeal adds information only when the replacement review can change the evidence or reasoning path. Recounting correlated votes changes the number, not the basis of judgment.
Agreement and language overlap answer different questions
Read the explanation
The available interface asks for three rulings and distinguishes outcome agreement from shared wording. Consider an explicitly illustrative example: two rulings agree and one differs. Two divided by three is about sixty six point seven percent. Three agreeing would give one hundred percent. The bars use three pixels per percentage point. This arithmetic illustrates a count and cannot prove a verdict correct or the reviewers independent. The local snapshot lacks its app JavaScript, so these example numbers are not claimed as actual tool outputs or verified implementation behavior. The interface labels its pairwise language metric Jaccard. As an explicitly illustrative definition, suppose one word set contains complete, work and approved, while another contains complete, work and rejected. The intersection contains two words, and the union four. Their ratio is one half. A two hundred forty pixel bar therefore represents the intersection on a one hundred twenty pixel per word scale, compared with a four hundred eighty pixel union. This does not establish how the unavailable app normalizes words, removes boilerplate, extracts outcomes or sets a caution threshold. Three rulings have three unordered pairs: A with B, A with C, and B with C. The interface provides pairwise language, obligations and a correlation caution separately from agreement. The bars compare three rulings with three pairs at eighty pixels per item. High shared language can arise from a common source, boilerplate or copying; this interface itself says language analysis is not a truth oracle. Without the referenced local scripts, current local browser checks can preserve the visible inputs, controls and layout but cannot prove live analysis or export. No legal outcome or real panel independence is certified here.
reasoning path, not the vote count.