A four-layer evidence pattern for apps that persist Diffusers outputs

[noqt]

We used Flux2KleinPipeline with the pinned black-forest-labs/FLUX.2-klein-4B revision f925183bed1a0440535d8190c1124d4b20a83696 in a local character-app prototype. The generation was only one layer of the evidence problem: the app then transformed, stored and later retrieved display files and a technical annotation.

A compact pattern made the result easier to inspect:

1. Generated scene — record model ID, pinned revision and pipeline.
2. Stored display files — publish exact file hashes and every transformation.
3. Stored observation — label application data as application data, not perception or human-like memory.
4. Restart retrieval — state exactly what the same persisted session returned after restart.

Worked example:
https://noqt.no-qt.chatgpt.site/research/zumi-one-session-two-frames-one-observation

The capsule shows two retained simulated phone-style frames, their SHA-256 hashes, and the bounded restart record. The public page is a static replay: it does not run the private app or model. The scene and frames are synthetic; no physical camera or person was used. The hidden scene plate is not published. The manifest records the retained JPEG hashes, lossless WebP display hashes, decoded-pixel roundtrip, pinned model revision and transformation details.

This establishes only what that retained record contains. It is not a claim about perception, memory, awareness, enjoyment, realism, human preference, adoption or novel model quality.

Sharing the pattern because it is reusable for Diffusers applications that persist outputs: generation provenance, stored files, application annotations and post-restart retrieval should remain separate claims.