Inspiration
A project folder is named FINAL. The contract is signed. The invoice is ready. Everything appears complete.
But what if a later amendment changes the agreed terms? What if an acceptance document is missing? What if the files change after someone approves the archive?
For small service businesses, closing a client project is not just a file-management task. It is a decision that depends on evidence scattered across contracts, invoices, amendments, and delivery records.
A folder can look complete without being ready to close.
FOLDERTRUTH was conceived to address that problem directly.
What it does
FOLDERTRUTH is an evidence-gated project closeout Skill designed for UGREEN AI NAS.
Instead of simply summarizing documents or organizing files, it aims to determine whether the available evidence actually supports closing a project folder.
The intended workflow is:
- Discover: Find relevant documents inside a user-authorized NAS folder.
- Reconcile: Compare contracts, amendments, invoices, and acceptance evidence against the user's explicit closeout criteria.
- Investigate: Search for additional evidence when documents conflict or important information is missing.
- Decide: Return
READY,BLOCKED, orINSUFFICIENT_EVIDENCE, supported by traceable sources. - Revalidate: Before a separately approved archive operation, check that the relevant source files have not changed.
- Verify: Re-observe the NAS state after an executed operation rather than relying on an AI-generated success message.
The default workflow is read-only. File modifications require explicit authorization.
The Winning Moment
Same request. Different evidence. Different decision.
A project initially appears ready for closure. A newly discovered governing document changes what the evidence supports.
FOLDERTRUTH is designed to revise its decision rather than blindly follow the original plan.
That adaptive behavior is the central hypothesis we intend to validate on UGREEN AI NAS.
How we built it
FOLDERTRUTH was developed as a Markdown-based Uliya Skill using the official UGREEN Skill Developer Kit 0.7.3.
Its orchestration design uses three registered expert agents:
file-agentfor file discovery, metadata, fingerprints, and authorized file operations.document-agentfor document extraction and reconciliation.knowledge-agentfor targeted evidence collection.
The Skill uses the supported delegation interface instead of custom NAS executables or direct filesystem access.
We also developed a separate local evaluation harness using synthetic project folders, document inconsistencies, and simulated file-state changes.
The final pre-test package contains SKILL.md and three supporting Markdown references covering evidence, safety, and validation.
Challenges we ran into
The biggest challenge was separating a convincing AI response from an independently verifiable outcome.
A document summary can sound correct even when it overlooks a governing amendment. Likewise, a successful file-operation response does not necessarily prove that the intended NAS state was achieved.
Another challenge was ensuring compatibility with Uliya's strict agent orchestration, permission handling, and Skill packaging requirements.
We also identified an important limitation: the documented filesystem capabilities do not establish an atomic compare-and-move transaction. Rechecking files reduces certain risks but cannot guarantee protection against every concurrent modification.
Accomplishments that we're proud of
The FOLDERTRUTH Skill passed the official Uliya 0.7.3 static validator with:
0 errors · 0 warnings · 0 diagnostics
We independently re-extracted the final ZIP and repeated the validation successfully, confirming that the submission package structure matches the checked Skill files.
We also established an evidence-oriented workflow with explicit uncertainty states, consent requirements, source references, and post-action verification steps.
These are design and static-validation accomplishments. We have not yet demonstrated the complete workflow on a physical UGREEN NAS.
What we learned
We learned that intelligent file management is not just about finding information. It is about knowing when the available information is sufficient to justify an action.
We also learned that a deterministic baseline can solve simple document-conflict examples. To demonstrate meaningful agentic intelligence, FOLDERTRUTH must handle previously unseen evidence, choose relevant follow-up observations, and revise its conclusions appropriately.
Most importantly, an AI's assertion of success is not the same as a verified outcome.
What's next for FOLDERTRUTH
If selected as a finalist, our next milestone is testing the Skill on the assigned UGREEN AI NAS environment.
We plan to evaluate consistent documents, conflicting records, missing evidence, and changed files after an initial review.
The critical tests will compare FOLDERTRUTH against a simpler deterministic workflow and independently verify the resulting NAS state.
Our goal is to turn FOLDERTRUTH into a reliable, privacy-conscious project closeout assistant that helps people make evidence-backed decisions before modifying important files.
