TrueBox

Inspiration

University hiring committees and talent-recruitment programs read hundreds of applications a year. Inflated claims are common and easy to miss: an academy membership dated a year early, a corresponding author written as first author, a filed patent written as granted. Checking by hand means cross-reading institution pages, paper databases and patent records, claim by claim. And these files are too sensitive to upload to a cloud AI.

What it does

TrueBox turns a UGREEN AI NAS into a verification desk that never goes online. The NAS holds three layers:

• a dated reference knowledge base (people, institutions with aliases, papers with author order, patents, awards),
• the organization's own private library (staff lists, past applications, review rules),
• the document under review.

The Skill extracts every checkable claim, anchors the person's identity (same name is not same person; when unsure, it asks), collects evidence only from the two knowledge folders, compares field by field, and writes a report: contradictions first, then every claim with its verdict, confidence and source. "Not found" never means "false". Reports are de-identified for blind review by default, and more deeply for public sharing.

How I built it

With the UGOS Skill Creation Tool. The Skill is a seven-step orchestration that delegates to the NAS's file, document and knowledge experts, passes only real results between steps, and handles every delegation state. The rules for extraction, identity, comparison, scoring and de-identification live in linked reference documents. It passes the official static checker with no errors or warnings.

Challenges

Writing rules precise enough that two different runs reach the same verdict. I built a test set (a reference snapshot of real researcher records, a fictional university talent office, and three documents with an answer key) and ran three rounds of blind, off-device simulations in which an AI agent played the Skill and the experts. Round one surfaced more than fifty unclear rules; round two found a real bug (two reports saved under the same name, one overwriting the other).

Accomplishments

In the last round: all five planted errors caught with sources, no false contradiction on a clean CV, and the Skill stopped to ask when a common name matched two people. The runs also flagged problems in the reference data itself, so verification doubles as a quality check on the knowledge base.

What I learned

Replacing a name is not de-identification: a career path or a list of awards can point to one person. The public level now shows awards as counts and keeps checksums and record IDs in a separate name key.

What's next

Test the rules with the NAS's built-in model on the real device, measure run time, and add on-device ingestion so new documents are structured and linked to the knowledge base as they arrive.

Built With

• ai-nas
• claude
• local-ai
• markdown
• ugos
• ugreen
• uliya-skill