TenetBench release package: a clause-anchored Russian corpus, generated tasks, judge verdicts and human annotations

Release package accompanying the paper "TenetBench: A Clause-Anchored Russian Corpus
for Evaluating Language Models against Professional Codes of Ethics".

Contents, 142 files in one archive:
- the corpus in reference mode: the normative clause text is withheld and published as
  a SHA-256 digest with a stated normalisation rule, together with locators and all
  derived fields; one document, an act of a state body, is released in full as a
  worked example;
- 4,470 generated task records (440 held out, 1,956 belonging to documents released by
  reference only);
- 7,303 judge verdicts over twenty panel rows;
- 4,385 human annotation rows over 1,260 items by three annotators and one
  adjudicator, under opaque identifiers, with free-text fields withheld;
- the generation and evaluation pipeline, the frozen judge prompts, the panel
  declaration and the release manifest.

Licences: data CC BY 4.0, code Apache-2.0 (LICENSE, LICENSE-DATA.txt,
LICENSE-CODE.txt inside the archive).

Integrity: archive SHA-256 249a7260d27aff4c...; content digest, taken over file names and bytes with container metadata excluded, c221e63b42f04e3b...