Inspiration
Industrial supplier reviews often combine quotations, specifications and certificates. A polished summary can hide mismatched models, test conditions or missing evidence. EvidenceGate is designed to make those limits visible before a human procurement decision.
What it does
Given a user-authorized document set and project requirements, the Skill orchestrates a source-linked evidence register: documented claims, conflicting evidence, missing evidence, incomparable conditions and unresolved coverage. An optional baseline review explains what changed and which questions remain open. It supports human diligence; it does not certify suppliers or provide electrical design approval.
How it is built
The Round 1 package uses the official UGOS Skill Creation Tool and its supported expert mappings: ug-filesystem through file-agent, ug-document through document-agent, and optional ug-knowledge processing or approved saving through knowledge-agent. It contains SKILL.md and directly linked Markdown references, with no custom executables or external API integration.
What is implemented
The original orchestration instructions, evidence comparison policy, synthetic acceptance cases and development disclosure are authored. The official static checker reports static-pass, with zero errors and zero warnings. This is not proof of installation or NAS runtime behavior.
Challenges and lessons
Local file storage does not establish that inference stays local. The workflow requires checking actual platform routing before confidential use. Document extraction completeness, source coverage, verified saving and search-index readiness are treated separately. Missing evidence must remain missing rather than becoming a confident answer.
Validation and next steps
If selected, validate on the assigned NAS: conflicting voltage claims, efficiency at an unspecified load, component-only certification, baseline changes, unreadable pages and embedded prompt injection. Record actual expert results and permission handling. No real supplier dossier, quote, contact data or credentials are included.
AI assistance
OpenAI Codex assisted research, workflow design, writing, packaging and static checks. Target-device execution and model routing have not yet been tested. The entrant must review and understand the work before final submission.
Built With
- ugos
