DriftBack – Instrument Lookback for UGREEN AI NAS

Inspiration

When a balance, thermometer, pipette or gauge is found reading wrongly, fixing it is the easy part. The hard part is the question every quality system asks next: which earlier results could it have affected? Standards such as ISO/IEC 17025 expect labs to examine the effect on previous results, but the evidence is scattered across scanned certificates, logbooks, test reports and spreadsheets from different years, and the same instrument is written as "BAL-07", "Bal 07" or just a serial number. Teams spend days on this hunt and still miss records. Calibration tools answer "when is it due?", not "what did this failure affect?"

A NAS is exactly where these years of mixed evidence live, so we built a Skill that answers that question where the data already is, without anything leaving the device.

What it does

DriftBack has two modes, both working only inside a folder the user authorises:

Lookback investigation: the user names the failed instrument (or just points to the failing certificate). DriftBack rebuilds its calibration history, bounds the period of concern, finds every record that used it, and ranks each result High / Medium / Low by how close it came to its acceptance limit, with the arithmetic and source file shown for every number.
Readiness scan: before any incident, it builds an Instrument Register that flags missing usage logs, inconsistent IDs, certificates linked to no work, overdue calibrations and drift trends, and suggests a lookback for any new out-of-tolerance certificate.

After the user reviews and approves, it saves an action pack (timeline, ranked records, evidence gaps, draft notification list, draft corrective-action report) and a dated register that the next run reuses.

How we built it

DriftBack is a pure orchestration Skill: a SKILL.md plus six Markdown references, with no scripts. Each step delegates to a registered UGREEN expert agent:

Step	Capability → agent
Inventory folder, load register	`ug-filesystem` → `file-agent`
Read certificates, results, limits, maintenance events	`ug-document` → `document-agent`
OCR scans, logbook photos, nameplates	`ug-photo` → `photo-agent`
Find mentions, transcribe voice notes, save output	`ug-knowledge` → `knowledge-agent`

The risk rule is explicit. For a result R and an as-found error E given as a percentage:

s = |R| × |E| / 100, and toward an upper limit U: C = R + s

High if \( C \ge U \), Medium if \( C + s \ge U \), otherwise Low. Example: \( 99.4 \times 1.008 = 100.2\,\text{g} \ge 100\,\text{g} \) gives High. The Skill also shows a direction-aware estimate, because a balance that reads high puts results near a lower limit most at risk.

The Skill passes the official Uliya static checker with 0 errors and 0 warnings.

Challenges we ran into

Designing only with capabilities that really exist. Our first design stored the register in a knowledge base and triggered itself whenever a new certificate was saved. The capability catalog supports neither, so the register became an approved, dated Markdown file that later runs read back, and automatic triggers became lookback suggestions from the readiness scan.
Avoiding false alarms. Look-alike instruments (BAL-07 vs BAL-09) had to be excluded, not matched. We defined four link levels (Confirmed, Probable, Uncertain, Excluded) and only rank the first two.
Getting the measurement logic right. The period of concern starts at the last calibration that proves the instrument was good. A percentage error on a °C reading is rejected, because Celsius has no true zero.
Never inventing data. Missing or unreadable values become "Manual review needed", and DriftBack never declares a result valid or invalid.

What we learned

The judgement in a calibration lookback lives in the time window and in each result's margin to its limit, not in matching IDs. Writing those rules out explicitly, with the arithmetic shown, is what makes an AI result something a quality manager can trust and audit.

What's next for DriftBack

If selected as a finalist: build a synthetic lab archive with planted aliases, a scanned certificate, a photographed logbook page, a log gap and near-limit results, and validate the full flow on a UGREEN AI NAS.