[hr-jin]

Summary

Task packages: tasks/financial-report-qa and tasks/visual-document-qa.
Author: Haoran Jin. Both existing tasks retain their 8-CPU, 32-GiB, CPU-only configuration and version 1.0.0.

• Financial-report evidence retrieval: valid repeated table rows were missing from the reference locations when table headers used different date formats, equivalent column names or separate text blocks. Add 18 reviewed alternative passages across nine reports, contributing 26 label locations to 18 held-out questions. Each addition was checked against the original row for reporting entity, financial item, period, unit and column meaning. Existing locations and evidence units are retained.
• Visual document QA: MMLongBench-Doc text, list and unanswerable answers previously went through extraction and text comparators, which could reject a correct complete sentence. Judge the full response semantically, with a binary score and reason. Keep integer/float extraction and comparison, the FinRAGBench-V judge, and the existing OpenRouter/Gemini judge configuration.

Instructions, public examples, reference answers, resource limits and score mappings are unchanged. Both task READMEs and provenance documents describe the corrections dated 2026-10-28.

Validation evidence

Run from the repository root with Python 3.12:

unshare -n python -m unittest discover -s tasks/financial-report-qa/tests -p 'test_*.py'
unshare -n python -m unittest discover -s tasks/visual-document-qa/tests -p 'test_*.py'
python scripts/check_tasks.py tasks/financial-report-qa
python scripts/check_tasks.py tasks/visual-document-qa
python scripts/check_release.py
python scripts/check_tasks.py --merge-ready --base bb0ddbaf105c25cc53e3598c37e0428c943a7098
git diff --check

• Financial tests: 35 passed, including actual corrected labels, duplicate alternatives, incomplete coverage and wrong-report rejection.
• Visual tests: 45 passed. All API responses are mocked and network access is disabled. They cover semantic routing, full-answer forwarding, response validation, retries, error propagation and the retained numeric/FinRAG routes.
• Corpus audit: all 18 new slices equal their corresponding original table rows; the 4,236 hidden questions and 6,603 hidden evidence units are unchanged. The five public examples are unchanged.
• Both task static checks, release checks, merge-ready checks and whitespace checks passed.

Task-local records: financial validation, visual validation.

Pending by request: after merge and pull, maintainers will re-evaluate the existing DeepSeek v4.1 Flash, GLM 5.3 Flash, GPT-6 Luna and MiMo v2.6 Flash outputs for both tasks. Financial scores can be recomputed from saved spans. Visual re-evaluation applies the new judge to the 626 MMLongBench-Doc Str/List/None questions and recomputes the aggregate score. No coding-agent trial, live judge call, submission re-evaluation or container rebuild was run for this change.

Contributor checklist

• Existing task authors and original commits preserved.
• Allow edits from maintainers.
• No secrets, downloaded corpora/models, job outputs or author-only solutions included.
• Static checks and offline regression evidence recorded; deferred runtime checks listed above.
• Provenance updated; corpus and immutable asset pins preserved.

Maintainer merge checklist

• Review the added evidence locations and semantic-judge behavior.
• Confirm the deferred re-evaluation plan and review CI.
• Merge while preserving the original contributor history.