Reconstruction evidence layer (AI-assisted provenance metadata)

The reconstruction "evidence layer" is an optional, AI-assisted way to record the provenance of each rebuild stage of a lost-place reconstruction. For any stage that has a caption (and optionally a body and a date label), a curator can ask the AI to suggest structured provenance metadata, review and edit it, and save it. The saved metadata then appears read-only to visitors in the public montage.

It is entirely optional and additive: a reconstruction and its montage work exactly as before when no evidence metadata is set.

What the evidence layer records

For each rebuild stage, the evidence layer captures four facets plus a short note:

• Date estimate - a human date or range, e.g. c. 1905 or early 1900s.
• Evidence type - photograph, survey plan, architectural drawing, written account, oral history, archaeological, comparable structure, inference, or unknown.
• Confidence - high, medium or low.
• Source credibility - primary, secondary, tertiary, conjectural, or unknown.
• Rationale - one short sentence explaining the assessment.

How a curator uses it

1. Open Reconstructions (admin) and find the reconstruction whose stages you want to annotate: /exhibition-space/reconstructions/manage.
2. Each rebuild stage has an Evidence layer panel (click to expand it).
3. Click Suggest with AI. The system reads the stage's caption and body text and proposes a date estimate, an evidence type, a confidence level, a source-credibility judgement, and a one-line rationale.
4. Review and edit every field. The AI only suggests; you decide what is correct.
5. Click Save evidence metadata. Clearing every field and saving removes the metadata again.

How visitors see it

On the public reconstruction montage (/reconstructions/{id}), open Show the rebuild stages as a list. Any stage with saved evidence metadata shows its date estimate, evidence type, confidence and source credibility as small badges, with the rationale beneath. Stages without metadata show nothing extra.

How the AI is called

The suggestion is produced through the AHG AI gateway (ai.theahg.co.za) using the platform's standard LLM service - never a direct model endpoint. The annotator reads the stage's caption and body text (not the image bytes), which is normally the richest cue for provenance.

The feature fails softly: if the AI gateway is unavailable, over quota, or returns an unreadable answer, the Suggest with AI button shows a short inline message and you can fill the fields in by hand. It never blocks adding or editing a stage.

Interpretive caveat

A reconstruction is one informed reading of the evidence, assembled for interpretation. The evidence layer makes that reading transparent - it records how confident the curator is and what kind of source each stage rests on. It is not a claim that the reconstruction matches the original's exact appearance.