AI Transparency Notice

When you run a scan, your photograph and the expected inventory for that rack are sent to a large multimodal model. The model — not a person and not a deterministic scanner — produces the list of detected assets, the per-detection confidence values, and the accompanying notes. Reconciliation against your records is then performed deterministically in our own code.

• In-app: every scan page and every scan result carries a persistent AI-generated notice.
• Exports: each CSV begins with a provenance header including the literal line THIS FILE CONTAINS AI-GENERATED CONTENT, the model identifier, the prompt version, and the generation timestamp.
• Machine-readable: exports are accompanied by a JSON provenance manifest describing the generator, model, and review state, so the marking travels with the file.
• Unreviewed exports are additionally watermarked DRAFT — UNREVIEWED AI OUTPUT.

These measures are intended to meet transparency and labelling obligations for synthetically generated information, including Article 50 of the EU AI Act (applicable from 2 August 2026) and comparable synthetic-content labelling regimes.

• The model can misread characters on worn, angled, reflective, or poorly lit labels.
• The model can fail to report an asset that is physically present but obscured.
• The model can report an asset that is not present, or transcribe a plausible but incorrect identifier.
• A 'no drift detected' result therefore means the model found nothing inconsistent — not that the rack is verified correct.
• Accuracy varies by lighting, framing, label condition, and hardware type. We do not publish a single accuracy figure because a single figure would not be meaningful across those conditions.

Every scan is provisional until a person reviews it. Reviewers can confirm, override, or reject each finding, and an override requires a written reason. The reviewer's identity, decision, reason, and timestamp are recorded and travel with the export. A scan that has not been reviewed is exportable only as a clearly marked draft.

If you believe a finding is wrong, record an override with a reason on the scan detail page. That record is retained as part of the audit trail. If you believe the system is systematically wrong, contact privacy@veracrat.com.

For each scan we store the model identifier, the prompt version, the generation timestamp, the structured findings, and the review decisions, so that any historical result can be reconstructed and explained.