Inspiration
A project handoff can contain two different deadlines for the same deliverable. A fluent summary may conceal that disagreement. Decision Ledger was designed around a more useful question: what does the evidence actually say, and what still needs a human decision?
What it does
Decision Ledger is a Uliya Skill for authorized UGREEN AI NAS project folders. Its workflow turns explicit source statements into documented commitments, two-source conflict dossiers, open questions, and evidence coverage. A human review queue gives each unresolved dispute a precise clarification question.
When the user explicitly supplies an earlier ledger, the workflow compares historical and current evidence. It distinguishes an incompatible claim from an explicitly documented replacement. Missing current evidence never means a previous commitment was cancelled. An optional meeting recording contributes provisional transcript evidence. Saving the displayed handoff requires approval of the content and destination, with separate overwrite consent.
How it is built
The package contains SKILL.md and one directly linked Markdown contract. It uses UGREEN's official uliya-skill-kit 0.7.3 and its published expert mappings: ug-filesystem to file-agent for discovery; ug-document to document-agent for extraction and reconciliation; ug-knowledge to knowledge-agent for optional transcription and approved text saving. Parent reasoning compares actual returned claims and any supplied baseline. No custom executable, database or background service is included.
The workflow limits analysis to 60 documents and one recording. It treats source content as untrusted data, preserves real provenance, stops dependent steps when evidence is absent, and discloses incomplete coverage. It requires processing-route confirmation before reading private content and prohibits external processing of sensitive NAS data.
Challenges and lessons
Different dates can describe different milestones. Multiple owners may be shared rather than incompatible. A later status can represent a legitimate change. The claim contract therefore matches subject and context before identifying a conflict, and requires explicit evidence for supersession. Another important distinction is that a successful static check does not establish runtime accuracy or privacy enforcement.
Current progress
The enhanced orchestration instructions and evidence contract are implemented. The official static checker passes with zero errors and warnings. Twenty-one local artifact and catalog checks passed. Sixteen synthetic validation scenarios and a finalist testing plan are prepared. No NAS execution or behavioral test pass is claimed: installation, extraction accuracy, history reasoning, transcription, permissions, model routing and persistence remain pending actual NAS validation.
What's next
On the assigned finalist NAS day, test conflicting deadlines, explicit replacement history, agreement, unreadable files, prompt injection, scope boundaries and approved persistence. Record actual results and demonstration footage, inspect every citation, and back up evidence before access expires.
Development disclosure
OpenAI Codex assisted with research, workflow and proposal drafting, synthetic test design and static verification. UGREEN's official developer kit supplied the capability catalog and checker. No third-party runtime libraries or assets are bundled in the Skill package. The core workflow was newly developed for this hackathon.
Log in or sign up for Devpost to join the conversation.