EvidenceDock — When a Decision Needs Proof, Not Just a Summary
The problem: AI can summarize a document. But can you defend the decision?
Picture a founder preparing to sign a supplier contract. One folder contains the latest proposal, an older contract, meeting notes, and a spreadsheet. A chatbot produces a confident summary: “The supplier agreed to a 30-day payment term.”
But an older version says 60 days. A later note mentions a possible exception. Where did the answer come from? Which source is current? What would a careful reviewer need to check before signing?
This is the problem EvidenceDock is designed to address.
Meet EvidenceDock: Source-Backed Decision Packets
EvidenceDock is a Round 1 pre-test AI Skill specification for UGREEN AI NAS, entered in the Automate Your Workflow track. Its proposed job is to transform authorized files into a decision packet that includes traceable evidence, conflicting claims, unresolved questions, and next actions.
Core principle: An AI-generated conclusion without inspectable source evidence is a suggestion, not a verified finding.
The difference from a conventional AI document summarizer is deliberate: EvidenceDock should resist supplying a clean answer when the sources do not support one.
The User Journey: From Scattered Files to a Reviewable Decision
- Ask a consequential question. Example: “What have we actually agreed to with this supplier?”
- Choose an authorized NAS folder. EvidenceDock checks the requested scope and permissions rather than assuming access to the entire NAS.
- Retrieve attributed passages. Its proposed workflow uses officially supported NAS file/document/knowledge capabilities, preserving source identity and version metadata where available.
- Compare the evidence. Conflicting versions, stale claims, missing sources, and incomplete support remain visible.
- Build a decision packet. The packet organizes findings, passage-level citations, contradictions, unknowns, and next actions.
- Verify and review. Claims without retrievable supporting passages are marked unverified or removed. The user previews the draft.
- Approve controlled export. Only after explicit permission does the workflow create a new file in an authorized destination. Original documents stay untouched.
Illustrative scenario, not a live NAS demo: Contract version A says payment is due in 60 days, while version B says 30 days. EvidenceDock should present both claims and their sources, identify the conflict, and recommend confirmation. It should not silently invent a reconciliation.
How I Designed It
The pre-test package follows the Markdown-based Uliya Skill structure:
- SKILL.md: User goal, inputs, capability boundary, expert-agent workflow, completion and failure states.
- references/workflow.md: Evidence contract, proposed output format, verification cases.
- references/safety.md: Authorized scope, document prompt-injection defense, privacy, consent and safe export.
- references/dependencies.md: Third-party components, AI-assistance disclosure and licensing boundary.
The planned architecture orchestrates officially documented registered expert agents and supported UGOS NAS capabilities, not arbitrary shell commands, custom scripts or direct filesystem access. Exact CLI names and interfaces must be checked against the official UGREEN Skill Developer Package before implementation; the draft does not invent them.
Inputs are the decision question, authorized file scope, approved output location, and consent. A supported local model would be preferred where available. Any optional external AI service would require explicit authorization and privacy review; no external model is presently integrated.
The designed failure paths include SCOPE_DENIED, UNSUPPORTED_INTERFACE, SOURCE_CHANGED, CITATION_MISMATCH and INSUFFICIENT_EVIDENCE. A failure is not relabelled as a successful export.
Why UGREEN AI NAS Matters
A NAS is where a team can already hold contracts, notes, reports and records. The proposed benefit is to reason over authorized, existing documents at their storage location, rather than treating an unconstrained chatbot prompt as the source of truth.
UGOS permissions, supported document/knowledge capabilities and authorized new-file export would each serve a concrete function in the workflow. Integration is pending validation, not claimed as commissioned.
Engineering Challenges and What Is Actually Complete
Scope before intelligence. A useful summary from an unauthorized folder is still a failure. EvidenceDock's specification places authorization before discovery.
Contradiction rather than false consensus. Different source versions can disagree. The workflow retains opposing evidence rather than averaging it into a plausible but unsupported answer.
Citation integrity. A model can fabricate a reference. Each proposed citation must be checked against a retrievable authorized passage.
Export as an explicit action. Drafting analysis must not silently modify a business record. The user authorizes creation of a new output.
Build status: The hackathon-period Round 1 Skill specification, workflow and safety references, and proposal are prepared. A verified UGOS CLI mapping, live NAS installation, actual extraction, on-device AI invocation, and operational export are not yet built or tested. This is a pre-test entry, not a disguised working demo.
How I Would Validate a Finalist Build
Core test 1 — Consistent sources: Three meeting notes agree on the same milestone; the resulting packet retains matching, inspectable citations.
Core test 2 — Conflicting versions: Two contracts contradict one another; the packet preserves both statements, identifies their versions, and refuses to manufacture certainty.
Edge test — Access or source changes: Permission is revoked, a source changes mid-process, or a file format is unsupported. The workflow stops with a clear reason and creates no unapproved output.
Finalists would map the Skill to the official UGOS tool interfaces and perform real tests during their assigned UGREEN AI NAS access period.
What I Learned
A useful agent must know not only how to answer, but also when evidence is too weak to answer. Source provenance, current permissions, and explicit uncertainty are product capabilities, not footnotes. The human decision-maker should own the final export and remain able to inspect what the system relied on.
What Comes Next
First, verify the official Skill interfaces and agent capabilities; next, implement authorized file discovery and attribution; then, test contradiction handling and citation checks; finally, validate permission gating, reviewed export and failure recovery on an assigned NAS.
Longer-term, the same workflow could help founders review commercial agreements, researchers assess literature, and teams turn scattered operational documents into defensible decision records. Those are future applications, not claimed customers or completed integrations.
Submission Materials and Disclosures
Round 1: The separate proposal PDF answers the organizer's ten questions within 500 English words. The pre-test ZIP contains SKILL.md and directly linked Markdown references. Neither a public repo nor a video is required in this round.
Principal hackathon work: The EvidenceDock Skill specification and core evidence-packet workflow were newly drafted starting November 7, 2026. Previously developed third-party Skills or submission materials are not reused.
AI assistance: OpenAI ChatGPT was used to draft and refine the description, Skill specification and validation scenarios. The official UGREEN UGOS Skill Developer Package is an external dependency awaiting compatibility review; no third-party code or credentials are bundled.
Truth boundary: EvidenceDock is currently a proposed workflow with prepared Round 1 Skill materials. It has not been demonstrated running on UGREEN AI NAS.
Because a decision is only as strong as the evidence we can actually show.
Built With
- ai-agent-orchestration
- document-analysis
- evidence-verification
- markdown
- ugos
- ugreen-ai-nas
- uliya-skill