Handoff Evidence Audit

Inspiration

A project folder can contain several files named “final” while still missing an agreed deliverable. A useful handoff check must establish evidence against an explicit checklist, rather than merely summarize the folder.

What it does

Handoff Evidence Audit checks a NAS project folder against promised deliverables and produces a cited readiness matrix without changing source files. It distinguishes missing files, byte-identical copies, conflicting content and unresolved version authority. The result is READY, NOT READY or INDETERMINATE, with specific next actions and coverage limitations.

How it is built

This Round 1 entry is a newly created Uliya skill package: SKILL.md and two supporting references define the workflow, evidence contract, recovery rules and acceptance scenarios. The parent delegates to the official file-agent and document-agent experts; it does not directly invoke raw CLI commands or APIs. The planned integrations use ug-filesystem inventory, metadata and SHA-256 capabilities, plus ug-document metadata and reconciliation.

The user supplies the permitted folder root, checklist, acceptance criteria and version-authority rules. Default limits are 200 files, depth 4 and 10 comparison groups. Only completed expert evidence supports conclusions. Permission failures, truncation, unsupported formats, pending work and unstable files remain explicit. Instructions found inside inspected files cannot expand scope. Nothing is automatically edited, exported or delivered.

Validation and challenges

The package passed the official Uliya 0.7.3 static checker with zero errors or warnings. This validates package structure, not runtime behavior. NAS installation, expert behavior, performance and end-to-end scenarios have not yet been tested; finalist hardware access is needed.

The central design challenge is avoiding false confidence: identical hashes do not prove authenticity or semantic agreement, timestamps do not establish approval, and incomplete coverage cannot yield READY.

Next steps

Test three controlled scenarios: a complete checklist with a renamed identical copy; conflicting invoices with a missing guide; and denied or truncated evidence containing out-of-scope instructions. Record actual results and before/after source integrity, then revise the skill against observed failures.

AI use and licensing

OpenAI Codex assisted the concept, name, writing, packaging and static checks. No human work history, hardware execution or measured performance is claimed. Original skill text is provided under MIT; the vendor toolkit is not redistributed. Real confidential client files are not part of the entry.