Inspiration
A NAS can contain thousands of contracts, invoices, scans and revisions. A matching filename does not prove that two files are identical. A convincing AI summary does not prove that its claims are true. We designed Evidence Ledger NAS to put traceable evidence before confident-sounding conclusions.
What it does
Evidence Ledger NAS is an evidence-first UGREEN Uliya Skill. Within a folder explicitly authorized by the user, it coordinates available NAS experts to:
- inventory document files and verify byte-equal duplicate candidates using real SHA-256 results;
- distinguish identical copies from possible revisions;
- extract supported document content and, when permitted and supported, read text from image files through native OCR/VLM;
- retrieve source-bounded semantic candidates and compare cited statements;
- produce a lightweight, sourced logical claim graph showing which document supports which statement;
- highlight possible conflicting claims, distinguishing overlapping effective dates, successive versions and unknown dates;
- issue a conservative review ledger that separates observed source text, OCR uncertainty, interpretation and missing evidence.
It is deliberately read-only by default. No files are deleted or modified, no people are accused of wrongdoing, and no authenticity or compliance is certified.
How we built it
The submission is an orchestration Skill, packaged as SKILL.md plus linked Markdown references, designed for the official UGREEN Uliya registered experts: file-agent (ug-filesystem and ug-filemgr), document-agent (ug-document metadata/extract/reconcile and native semantic_lookup), photo-agent (image-only vision-ocr/vision-vlm) and knowledge-agent (collect-evidence). Delegation is dependency-ordered, source-scoped and fail-closed. Report persistence requires the user's separate permission.
An independent standard-library Python research harness tests deterministic source-to-claim graph rules with synthetic supplied records; it is not included in the Skill ZIP and does not run on the NAS. We do not claim that the platform exposes custom embedding storage, a persistent graph database, Qwen/Ollama or Python execution in Skills.
Testing and what is verified
The official Uliya static Skill checker returned static-pass, zero errors and zero warnings. 15 offline unit tests passed on a separate research prototype, covering exact-hash candidates, date overlaps and disjoint revisions, unknown applicability, unsupported/malformed evidence and bounded graph review. The Round 1 PDF contains ten answers within the official 500-word ceiling. A physical or remotely provisioned UGREEN NAS runtime test has not been performed; validation on the actual platform is a Round 2 gate if selected.
Challenges and learning
We avoided a common evidence-analysis trap: conflating similar names, OCR text, model output, file metadata and verified facts. We report uncertainty and partial failures instead of inventing missing provenance. Two differing amounts can be successive revisions, not contradictions, if their validity periods do not overlap.
What's next
With authorized UGREEN NAS access, test actual expert delegation, image OCR, supported document formats, bounded semantic lookup, hash verification and read-only behavior against two reproducible core fixtures and one edge case. Any unsupported capability will remain an explicit limitation.
Dependencies and AI-use disclosure
Platform: official UGREEN Uliya Skill Developer Package and registered UGOS experts, governed by UGREEN developer terms. No third-party executable, hosted API, paid model key or model weights are bundled in the submitted ZIP. OpenAI ChatGPT/Codex assisted design, documentation and static verification. This is a skill proposal with validated structure, not proof of live NAS deployment.
Source code
https://github.com/issueslinux/hunter (research/private project; the actual candidate Skill is in the attached ZIP).
Built With
- nas
- ocr
- ugos
- ugreen
- uliya