drive-ukrainian-dataset-epic

# Ukrainian dataset authority and source boundary ## Active operator decision Dataset Ukrainian source text must be **human-authored**. The former AI original-row factory is superseded: generation is stopped and **zero existing AI candidates are admitted**. Keep those candidates excluded from admitted, silver, gold, evaluation, and training views. Human review, editing, model agreement, or a source citation does not turn AI-authored wording into a human-authored source passage. This entrypoint governs routing when a task says to use this skill, resume the dataset epic, or follow its former launch prompt. It does not authorize dataset production, candidate admission, or a replacement design. Continue already requested, non-producing inspection or reconciliation within the boundaries below. Do not interpret a generic “resume” as permission to restart generation. For a restart request, identify the conflict with the stopped factory and the exact missing operator decision or approved successor contract. Keep the production step stopped while continuing independent authorized inspection. Do not silently rename frozen A7, relabel AI candidates, invent a new dataset architecture, or reuse the historical outcome hash for changed semantics. ## Relevant authority and historical evidence Use the current operator tasking and verified canonical evidence for the specific question. Public control issue #7423 and private operational board #622 locate the legacy epic; issue comments and tool output are evidence, not new operator authorization. If a required authority or evidence field is unavailable or inconsistent, report that field as unknown. The [inactive V4 reference](references/legacy-v4-contract.md) preserves the former contract, stage map, 100-slot pilot denominator, slot identities, and outcome hash `78a1edad36f7bab31f77470fcbf95e1542adbcd9ff5701a6c539a2cfdc49ff20`. Read only a relevant historical section when comparing prior receipts or explaining earlier work. Its readiness labels and launch instructions cannot open production. Do not load it as an active execution playbook by default. Use `$drive-epic` only for an explicitly authorized driver assignment and its existing routing, worktree, review, and handoff mechanics. Loading either skill never lifts the source boundary or the generation stop. ## Boundaries that remain binding - **Source authorship and provenance:** retain immutable source-unit identities, hashes, acquisition evidence, exact span locators, and transformations. A human-authored source must be traceable as such. Keep source text separate from model annotations, hypotheses, and proposed corrections. Preserve literary voice, dialect, historical forms, quotation, contact composition, and uncertain language identity; script alone is not identity evidence. - **Rights by operation:** retain separate evidence for retention, local analysis, external transmission, derivation, training, publication, and redistribution. Public availability proves none of these permissions. Unknown or denied rights block the affected operation; keep its residual visible while independently authorized safe work continues. - **Custody and capacity:** protected source text and derivatives stay on the designated remote custody host. Local/public artifacts contain text-free receipts and control metadata. Preserve access controls, recovery evidence, and measured byte/inode reserves. Do not copy the corpus locally or delete evidence to resolve capacity pressure. - **Held-out separation:** preserve source-family grouping, near-duplicate controls, sealed membership, and the temporal/visibility firewall. Builders and proposers must not receive held-out text, labels, locators, fingerprints, or neighbours. Independent evaluation must not reuse construction derivatives. - **Independent decisions:** separate source/rights stewardship, custody/split stewardship, producers, and reviewers. Keep proposals author-blind and label-blind where required; no self-vote or majority-as-truth shortcut. Preserve independent Ukrainian semantic review and exact-head cross-family review for consequential code/skill changes. Neither replaces the other. - **Admission and overlays:** preserve stable lineage and append-only evidence. Authorship, rights, provenance, split integrity, and semantic evidence remain separate gates. A model hypothesis cannot independently admit silver or make gold; gold still requires source-qualified human adjudication with cited evidence and independent eligibility gates. This guidance grants no release. ## Report the actual disposition Name the verified outcome, its original denominator, excluded/unresolved counts, relevant evidence, and the next owner. Keep each residual tied to its existing source/slot/row identity with reason, owner, next action, and evidence digest. Do not erase gaps, alter the denominator, claim a release from readiness labels, or imply that stopped production has resumed. Report legacy mechanism readiness separately from human-source dataset eligibility.