Synthetic non-agentic case: Multimodal product-catalogue accessibility assistant

Fictional example. The company, catalogue, assets, evaluation results, and decision are synthetic. They illustrate a controlled multimodal workflow and are not evidence of product performance.

Dominant pattern: multimodal. Secondary patterns: generation and classification. Level: A1. The system reads authorized product images and packaging text, then drafts alt text and flags inconsistencies. It cannot edit the source asset, publish a page, or write to the catalogue.

1. Starting point

Asteria Home SA is a fictional Swiss home-goods SME with 2,400 products sold in Switzerland and the European Union. New catalogue entries arrive with product photos, packaging images, dimensions, material fields, and supplier text. An editor drafts French and English alt text and checks visible claims against the product record.

The target is: draft factual alt text and flag visible inconsistencies for an editor. The system does not infer quality, safety, origin, sustainability, or protected characteristics from an image.

2. Rights, provenance, and system boundary

Every asset has an owner, licence or supplier authority, source identifier, hash, capture date when available, and permitted-use record. Faces, homes, addresses, vehicle plates, customer uploads, and unrelated background people are excluded from the pilot.

The source asset is immutable. The system may not generate a replacement image, remove a watermark, strip provenance, or silently alter packaging text. If a future version generates or manipulates public media, the provider/deployer Article 50 analysis and machine-readable marking test reopen before use.

3. Multimodal evaluation

The frozen set contains 140 image sets: 20 low-resolution images, 20 text-heavy packages, 20 near-duplicate colours, 16 deliberate image-record mismatches, 12 unsupported sustainability claims, and transformation tests for crop, resize, compression, and metadata loss.

Synthetic frozen-set result

Most corrections concern colour nuance and overly long descriptions. Two small packaging warnings are missed after aggressive compression. Compression limits are therefore part of the input contract rather than hidden in average quality.

4. Shadow workflow

For four weeks, editors complete their normal work before viewing the proposal. They record acceptance, correction type, missed visible information, invented information, rights exception, mismatch decision, active review time, and final published alt text. The public catalogue receives only the editor-approved text.

The workflow stops on a rights failure, invented certification or safety claim, critical accessibility omission, unflagged deliberate mismatch, or provenance loss. New modalities, generated images, faces, or customer uploads require a new risk and legal gate.

5. Decision

Decision: authorize a 60-set shadow pilot for the defined product-photo scope.

The frozen set supports observation, not automatic publication. Any measured time change must be reported with correction rate and accessibility quality. A faster draft that shifts work to later correction does not pass the business gate.

6. Transfer limits and source anchors

The figures are synthetic and cannot be generalized to people, medical images, insurance evidence, biometric systems, video, or voice. Those uses have different rights, harms, and evaluation contracts.

The provenance and transparency prompts use NIST AI 100-4, C2PA 2.2, the Swiss authorities' statement on AI-generated images, and the AI use-pattern guide.

7. Evidence pack

Retain asset authority, source hashes, transformations, the frozen set, expected visible attributes, alt-text rubric, accessibility review, provenance tests, correction ledger, exclusions, incidents, and signed gate decision.