Inspiration
A NAS can store the evidence behind everyday decisions: invoices, receipts, delivery records, photos, documents, and notes. But when two records disagree, finding more files is not necessarily the answer.
I kept coming back to a simple question:
What if an AI didn't just search for relevant evidence, but could reason about which evidence would actually settle the disagreement?
That became RECON — Evidence Resolution Intelligence.
The idea was inspired by the gap between retrieval and reasoning. A search engine can find a document. RAG can summarize it. An AI can propose an explanation. But none of these necessarily answer the most important question:
What should I check next to distinguish between competing explanations?
RECON was built around that question.
What it does
RECON turns conflicting private NAS records into a bounded evidence investigation.
Instead of immediately searching for more information, it first creates competing explanations for the conflict.
For each explanation, RECON predicts what evidence should exist if that explanation were true.
It then searches the user's authorized NAS evidence and ranks possible follow-ups according to:
- Discrimination — how strongly the evidence separates competing explanations
- Relevance — how directly it addresses the unresolved question
- Reachability — whether the evidence is actually accessible
The result is a NEXT-BEST EVIDENCE decision rather than another generic search result.
The core flow is:
Discover → Extract → Detect Conflict → Form Hypotheses → Predict → Rank Evidence → Follow Up → Compare → Verdict
RECON also knows when it does not have enough evidence. Missing evidence is not silently converted into a confident conclusion; the system can return UNCERTAIN or INSUFFICIENT EVIDENCE.
How we built it
I built RECON as a UGREEN AI NAS Skill with a parent orchestration layer coordinating specialized capabilities.
The workflow uses:
- ug-filesystem / ug-filemgr for authorized record discovery
- ug-document for document and text evidence
- ug-photo for visual evidence
- ug-knowledge for stored notes and knowledge
The parent Skill coordinates these capabilities through bounded stages.
The most important design decision was to separate facts, hypotheses, predictions, and observed evidence.
RECON does not simply inspect a result and construct a story afterward. It first records what each hypothesis predicts, then performs the focused follow-up, and finally compares the prediction against the actual evidence.
This creates a reasoning loop:
Evidence
↓
Competing Hypotheses
↓
Predicted Evidence
↓
Evidence Ranking
↓
Focused Follow-up
↓
Prediction vs. Reality
↓
Sourced Verdict
The proposed core workflow is read-only and designed around user-authorized NAS data. No external runtime service is required for the core Skill.
Challenges we ran into
The hardest part was realizing that relevance is not the same as usefulness.
A document can be highly relevant to a question while providing almost no ability to distinguish between two explanations.
That forced me to rethink the workflow from:
“Find the most relevant file.”
to:
“Find the evidence that changes the decision.”
Another challenge was preventing the AI from becoming overconfident.
If evidence is missing, the easiest behavior for an AI system is to fill the gap with a plausible explanation. RECON instead treats uncertainty as a first-class outcome.
I also had to design the orchestration carefully so that specialized NAS capabilities contribute real evidence rather than becoming disconnected AI calls. Each stage therefore consumes structured outputs from the previous stage.
Finally, the biggest remaining challenge is real-device validation. The Skill has been designed and documented, but its behavior still needs to be tested against the actual UGREEN AI NAS runtime and capabilities.
Accomplishments that we're proud of
The biggest accomplishment is that RECON is not simply another AI search assistant.
Its central mechanism — prediction before evidence retrieval — gives the Skill a different reasoning model.
I built:
- A complete Skill specification
- Expert-agent orchestration
- Competing hypothesis generation
- Prediction mapping
- Next-best-evidence ranking
- Bounded follow-up logic
- Explicit verdict and uncertainty states
- Safety and permission boundaries
- Reference documentation
- A structured finalist testing plan
I am especially proud of the decision to make “insufficient evidence” a valid result.
A system that knows when it cannot justify an answer is more useful for private records than one that produces a confident answer every time.
What we learned
I learned that building an AI system around what to retrieve is fundamentally different from building one around what can be proven.
The most valuable evidence is not always the closest semantic match. Sometimes a seemingly secondary record is dramatically more informative because it separates two competing explanations.
I also learned that uncertainty should not be treated as a failure state.
For evidence-based AI, saying “the available records cannot distinguish these explanations” can be a better result than generating an answer that sounds convincing but cannot be supported.
Most importantly, I learned that the strongest AI workflows are not necessarily the ones that generate the most content.
They are the ones that make the next decision better.
What's next for RECON — Evidence Resolution Intelligence
The immediate next step is real-device validation on the UGREEN AI NAS.
I want to test RECON against realistic private-data scenarios involving invoices, orders, delivery records, documents, and photos, measuring whether it consistently chooses evidence that actually separates competing explanations.
From there, RECON could evolve from resolving individual discrepancies into a broader evidence reasoning layer for private NAS data.
The long-term vision is simple:
Don't just find the evidence. Find the evidence that can decide.
Built With
- agents
- artificial
- decision
- document
- evidence
- explainable
- hypothesis
- information
- intelligence
- investigation
- kit
- knowledge
- management
- nas
- reasoning
- retrieval
- skill
- testing
- ugreen
- uliya

