ClearDesk

Motivation

ClearDesk explores whether a local model can select an exact supporting passage instead of producing an unsupported procedural answer.

What it does

Apertus selects a source and quotation from four synthetic IT procedures. A deterministic gate verifies the source ID and verbatim quotation. Unsupported questions yield abstention. The application takes no actions and has no live institutional integration.

Implementation

Built from the official HackApertus template using Apertus 1.5, Node.js, Docker Compose and GNU Make. Model weights are obtained separately under upstream conditions and are not redistributed. AI-assisted development is disclosed.

Verification and challenges

Root make run completed twice with exit 0. Actual runtime inspection confirmed a read-only model mount, no published host ports, an internal network and cleanup. Contract tests passed 22/22. Seven authored development cases passed 7/7 in the final container run; an earlier failed 5/7 baseline is preserved. Improving the constrained output and deterministic quotation gate addressed the observed baseline failures.

Limitations

These small development-visible fixtures are not held-out accuracy, real-world reliability or security certification. The corpus is synthetic, not a university or employer policy. The 70-second demo is playback of actual generated output, not a live Docker inference recording.

Reproducibility

The public template-derived repository includes source, tests, fixture provenance, full license notices, runtime evidence and a three-page technical report. Obtain the specified model separately, accept its upstream conditions and run make run from the repository root. No API key is required.

Built With

• node.js