All notable changes to ZxBench are documented here.

• Restore all 20 pre-existing multi-file long tasks to formal scope: benchmark 1.31.1 runs 620 questions, including 150 programming questions.
• Make benchmark.json the only released import catalogue. Development fixtures and historical CR2 subsets are no longer imported as formal questions.
• On upgrade, retire only accidentally imported bundled-history rows while preserving user-created custom questions and historical result rows.
• Bind official-run selection and dashboard counts to the released catalogue, and reject same-ID database definitions whose frozen content hash is out of sync.

• Restore the released 正式评测 flow: the 580 frozen, score-bearing public scenarios are now accepted; 20 explicitly marked multi-file development-shadow scenarios remain excluded from the main score.
• Normalize goldVerifiedAt before scenario hashing, so SQLite/Prisma date round-trips cannot falsely report a frozen definition as modified.
• Synchronize the local 600-scenario definition store without altering historical runs or result rows, and add API/contract regression coverage for the released flow.

• Publish benchmark 1.30.0: 600 active definitions across 10 dimensions, with 78 retired scenarios retained as versioned archive material.
• Add five prospective-only, deterministic challenge questions for hallucination resistance and mathematical reasoning. They use strict JSON contracts and no Judge calls.
• Complete the trusted single-file programming migration: 107 repair scenarios, 345 frozen verification IDs, and two executable-evidence PR scenarios.
• Expand and freeze the data-extraction suite at 56 typed JSON-contract scenarios.
• Repair real-time monitor card contrast in dark mode.

• Multi-file project_repair scenarios remain explicit-only development shadow tasks. JavaScript, Python, and Go maintainer risk samples have replayed positive gold; C#, Rust, and SQL remain pending and do not affect official scores.
• MC2-004-R1 is a redesigned development candidate, not part of benchmark 1.30.0 or historical score comparisons.
• Complete fenced JSON is no longer misclassified as truncated; genuinely unclosed fences remain rejected.

• Automated unit and contract regression tests cover the release changes. Container-backed gold replay verifies the initial JavaScript, Python, and Go multi-file risk sample.