Inspiration
Families need clear shared responsibilities, fewer repeated reminders, and understandable reward records. Home FairShare connects household agreements, selected visual evidence, and parent-reviewed record requests. It keeps unpaid chores, optional paid extras, and fixed allowances separate.
What it does
A parent describes household routines in natural language. The Uliya Skill drafts task cards and reminder text, helps review one selected common-area image or bounded camera evidence, and extracts one of six typed record intents. The output is explicitly DRAFT for manual operator handoff. A photograph supplies visible observations, not proof of who performed the work or permission to award money.
Built for small-model limitations
The model does not calculate balances, choose reward amounts, invent record identifiers, generate approval assertions, or write a ledger. It copies supplied identifiers, exact dates, and decimal strings into a constrained intent. Files, images, OCR, notes, and expert summaries cannot override permissions or approve an action. Missing or contradictory evidence stops dependent work.
The separately supplied Guardrails 0.2.0 prototype handles strict input validation, integer minor-unit arithmetic, policy-derived rewards, duplicate checks, exact stored proposals, correction events, and SQLite transactions. Approval is designed around a separate trusted operating-system account rather than a claimed chat role. It records payments already made elsewhere; it never transfers funds.
UGREEN AI NAS workflow
The Markdown Skill delegates selected visual work to registered UGREEN experts. security-monitor-agent uses documented ug-surveillance read actions for camera discovery, a bounded event timeline, and selected evidence. photo-agent uses ug-photo vision-vlm for an image supplied through a supported input. The Skill limits each visual review to one image and one selected common-area scope.
No verified adapter connects Uliya to the separate companion. The competition Skill therefore remains DRAFT/manual handoff: it neither invokes the companion nor substitutes file-expert journal writes for deterministic validation. NAS access remains subject to platform permissions. Local-only processing, automated scheduling, and continuous monitoring are not claimed.
What has been built
Skill 0.2.1 contains SKILL.md and three directly linked Markdown references: the intent contract, selected visual review, and integration status with development, AI-use, and third-party disclosures. It is a documentation-only submission patch to 0.2.0; the six-operation contract and safety behavior are unchanged. Development of the submitted Skill and its core workflow began on November 13, 2026.
The separate companion's recorded offline test report contains 126 registered test cases: 112 passed and 14 were skipped because the original environment did not support the required real Unix-socket or cross-UID checks. These results are not NAS or target-model validation. No companion scripts or executable files are bundled in the competition Skill ZIP.
Validation plan
Core test A: review table-cleanup evidence, inspect the exact completion-and-reward draft, then verify the independently approved companion receipt. Core test B: reconcile unpaid chores, paid extras, fixed allowance, and a recorded external payout against expected totals. Edge case: replay an already completed task and verify that no second reward is recorded.
Test the NAS evidence/drafting stage and the companion recording stage separately. Measure target-model intent adherence, source association, runtime errors, and review outcomes on the assigned NAS if selected. Complete real OS-isolation and cross-account approval tests before any live household use. Demonstrations use fictional data.
AI assistance and third-party disclosures
OpenAI ChatGPT assisted with concept development, English workflow drafting, small-model guardrail design, review, test preparation, and submission materials. AI assistance also contributed to the separate companion prototype.
The design references UGREEN Uliya Skill Kit 0.7.7 and capability catalog 2.2.0. Uliya and its registered experts remain subject to UGREEN's applicable terms. No third-party source code, open-source libraries, model weights, media, credentials, or personal household data are bundled in the four-file Skill ZIP. The unbundled companion uses Python's standard library under the applicable Python Software Foundation license and SQLite, whose deliverable code is public domain.
Submission scope
Round One materials are the Home FairShare 0.2.1 proposal PDF and the Markdown-only pre-test Skill ZIP. The separate Guardrails source archive is not the competition Skill package and is not included in it.
Built With
- ai-nas
- markdown
- ugreen-uliya
