# Thinking, design and gaps in an archive

Two small work samples prepared for Andreas Ehstand with AI assistance. The instruction score presents one text in two layouts. The archive template uses six synthetic messages to show how available context and missing material determine which conversation pairs are included. The materials are in German. Editorial date: 7 September 2026.

## One instruction score in two layouts

“Der Zwischenraum” contains 21 German words. Both layouts preserve every word, punctuation mark and the order of the text. Typeface, alignment and spacing differ; longer sentences wrap within their paragraphs. No effect on readers has been measured.

- [Readable HTML version with an editorial and source note](f09_denkpartitur.html)
- [PDF spread: two layouts on one A4 landscape sheet](f09_denkpartitur_doppelseite.pdf)
- [Plain score text](f09_denkpartitur.txt)

The score was newly formulated for this sample with AI assistance. The HTML page links to a [MoMA collection entry](https://www.moma.org/collection/works/131347) as a historical comparison for the relationship between an instruction and a particular realization. An artwork's date, a version's production date and a museum label's date refer to different things. The editorial date above dates this sample.

## Archive template with a filled example

The filled example shows how original text, later summaries, analysis and missing material are documented separately. Its six message texts are copied unchanged from a pre-existing synthetic work sample. Under the specified single-turn rule, three user-answer pairs are eligible, sharing two distinct user contexts; one answer is excluded because its parent message is missing.

- [Filled template with six messages in Markdown](f11_ausfuellbeispiel.md)
- [The same filled template in JSON](f11_ausfuellbeispiel.json)
- [Public source excerpt containing all six message texts](f11_quellenausschnitt.json)
- [Python validation script](f11_validieren.py)
- [Saved validation result in JSON](f11_pruefergebnis.json)

The six available messages contain 34 words in total. The five distinct eligible messages contain 30 words; summed across the three pairs, the total is 35 because one user message is counted in two pairs. These counts describe only the selected synthetic texts. No model groups can be established without documented model provenance. Including a pair does not establish that its answer is true.

The blank template remains available for use. Its entry fields and record lists are intentionally empty. Original, export and analysis dates have separate fields. Copy one record block for each message and one gap block for each missing item. The earlier single coding case presents the same excluded answer separately.

- [Blank archive and gap template in Markdown](f11_archivvorlage.md)
- [Blank archive and gap template in JSON](f11_archivvorlage.json)
- [Earlier single synthetic coding case in Markdown](f11_kodierbeispiel.md)
- [The same single coding case in JSON](f11_kodierbeispiel.json)
- [German notes on the archive sample](f11_liesmich.md)

The single coding case contains one invented answer with four words. Its referenced parent message is absent, leaving zero eligible user-answer pairs in that case. The generating model and version remain unknown in both examples. A role label does not establish this provenance.

The examples contain no private conversations and do not constitute an archive investigation, a user study or a new model experiment. The [OASST1 dataset card](https://huggingface.co/datasets/OpenAssistant/oasst1) is a structural reference; the message texts were not copied from that dataset. A file hash supports a comparison of file contents but does not establish a historical publication date or invention priority on its own.

## Further reading

The published German article [Checking whether AI ideas hold up](https://augmanitai.com/fachbeitraege/ki-ideen-tragfaehigkeit-pruefen/) complements these samples with a discussion of examining justifications.

[Deutscher Lesehinweis](README_DE.md)