Inspiration
When something goes wrong, such as a damaged deposit, a denied insurance claim or a faulty supplier delivery, the evidence is usually a pile of phone photos, receipt snapshots, a quote and an invoice. The other side then asks the same questions: is it complete, do the dates line up, do the amounts match, and has anything been edited since?
Cloud AI tools can summarise a pile like this, but they mean uploading private photos and paperwork, and a language model may invent a date or a total. A NAS already holds the files, keeps them private and has local AI, so it is the natural place to build evidence you can trust.
What it does
ProofPack is a pure-Markdown Uliya Skill with two modes.
Build mode. You point it at one case folder. It:
- inventories the files
- computes a SHA-256 checksum for each
- reads photo capture metadata and visible text
- extracts fields from receipts, quotes and invoices
- checks quote totals against invoice totals
- builds a dated timeline
- raises neutral flags, such as a missing capture time or an amount difference
It then saves one Markdown evidence pack, after platform approval. Every line is labelled NAS-verified, AI-observed or User-reported, so a reader can see what a tool measured and what a model described.
Verify mode. Later, you give it a saved pack. It re-hashes the files and reports which are unchanged, changed, missing or new.
Originals are never modified. The Skill makes no claims about fault, fraud or legal outcome, and it says plainly that a matching hash shows a file is unchanged since sealing, not that it is authentic.
How I built it
The Skill is only Markdown, as the Uliya standard requires: one SKILL.md plus five linked reference files (pack format, consistency checks, privacy and limits, disclosure, validation plan). The workflow is a fixed chain of delegated steps, and each step consumes only verified results from the one before:
- file-agent (
ug-filesystem,ug-filemgr): inventory, SHA-256, duplicate scan - photo-agent (
ug-photo): metadata, text recognition, one-sentence descriptions - document-agent (
ug-document): field extraction and amount reconciliation - knowledge-agent (
ug-knowledge): approved save with read-back
A design rule runs through all of it: a model never produces a path, hash, date or total. Those values are copied verbatim from NAS tool results. After a verified save, an automatic seal check re-hashes the files against the saved manifest, so a copying error in the saved pack is caught immediately.
I built it against the official Skill Creation Tool (kit 0.7.7, capability catalog 2.2.0) and ran its static checker until it reported zero errors and zero warnings.
Challenges
- No code allowed. With no scripts, all the reliability has to come from workflow design: explicit status handling, stop conditions, approval gates and honest limits.
- Trusting model-read hashes. A model compares the digests in the Skill, which is a real weak point. The seal check, full-digest display on any mismatch, and a planned test that hand-edits one digit in a saved manifest are the mitigations.
- Unknown runtime behaviour. I could not test on hardware before Round 1, so the proposal separates what is built from what needs NAS validation.
- Checker strictness. The checker rejected the kit's own template heading, and it bans certain bypass wording even in negated sentences, so I rewrote sections until the package passed.
- A 500-word proposal. I counted words from the rendered PDF to stay under the limit.
What I learned
- Provenance labelling makes AI output far easier to trust: a reader can see which lines came from the NAS and which from a model.
- On a Skill platform, the hard design problem is failure handling, not the happy path. Partial results, rejected approvals and empty folders each need a defined outcome.
- An integrity check is only credible if it states what it does not prove.
What's next
If selected, I'll test on the assigned iDX6011 Pro. The plan covers five functional tests, three edge cases, failure recovery (including a rejected save approval), one performance run and one security test. These include:
- manifest digests checked against independently computed values
- an edit, a deletion and an addition detected by verify mode
- a check that originals stay byte-identical after a build
Built with
Uliya Skill (Markdown), UGOS Skill Creation Tool 0.7.7, UGREEN AI NAS experts (file, photo, document, knowledge). No third-party libraries. Claude (Anthropic) assisted with research, drafting and static checking.
Built With
- ai-agents
- document-extraction
- local-ai
- markdown
- multi-agent-orchestration
- ocr
- on-device-ai
- photo-metadata
- privacy
- sha-256
- skill-md
- ugos
- ugreen-ai-nas
- uliya-skill