Inspiration

Imagine opening the dashboard the morning after a training run. Thirteen jobs finished, and there are thirteen scores to report. Then you notice that one job used a different learning rate. Only twelve belong in the comparison you planned.

That’s the question that led me to HELD. When I ask “Which runs count?”, I want to see the line in the plan and the line in the run record that explain the answer.

What it does

HELD is a Uliya Skill for researchers who keep experiment plans and run records on a UGREEN AI NAS. Its workflow produces a card showing how many planned seeds have eligible results, which settings changed and how the scores compare with the registered hypothesis.

In the learning-rate example, the card identifies group A, seed 7, and shows both values: the registered 0.001 and the observed 0.01. A researcher can inspect the source for each value and approve saving the card, a change summary, a failure ledger and a checkpoint.

How I built it

I wrote the Skill in Markdown using UGREEN’s Skill Creation Tool. The file expert reads the selected run records, the document expert extracts a plan from a PDF or Word document, and the knowledge expert can look up related papers locally. In the proposed workflow, Uliya processes the selected data on the NAS. Source folders are read-only. Saving to the output folder needs approval, and the workflow calls no external service.

Challenges I ran into

Five hours of silence tells me when a job last reported. It leaves the outcome open. I made that a warning and reserved failure rows for events with a recorded cause. I also had to decide what happens when a researcher revises the plan after a failure. The event keeps its original ledger row and plan hash. The next digest records any change in how the researcher interprets it.

What I’m proud of

The Round 1 package passed UGREEN’s static checker with zero errors and warnings. I also checked the expected counts with a separate local calculation. The synthetic baseline yielded 13 eligible seeds out of 20. Changing one completed run’s learning rate brought the count to 12. In a separate case, a five-hour silence left the count at 13 and produced no failure row.

What I learned

A count is easier to trust when I can inspect the one run that changed it. “A7 used 0.01. The plan specified 0.001.” Those two lines give a researcher a concrete reason to check the source or rerun that seed.

What’s next for HELD: Hypotheses, Experiments, Ledger, Drift

If selected as a finalist, I’ll compare the three cases on the assigned NAS with their expected results. I’ll also deny a save and check for zero writes, interrupt an append and retry it, then revise the plan and verify that the original failure still occupies one row. Finally, I’ll measure how long a 200-run snapshot takes.

Built With

  • built-with-ugreen-ai-nas
  • json
  • markdown
  • ugos
  • uliya
  • uliya-skill-kit
Share this project:

Updates

Submission history