I wanted to test GPT-6 Astra on a real task rather than a prepared demo. The challenge was to structure a real-world archive of 1,225 files across 97 folders, preserve existing numbering, assign missing numbers, identify duplicates without deleting them, and build a verifiable central index. The most interesting part was not just the final result, but what happened during the long-running process. Interruptions, conflicting progress states, temporary results that were no longer available after a restart, and eventually a maximum conversation length made persistent checkpoints essential. After introducing a persistent checkpoint, the workflow became much more reliable to resume. The project therefore became both an archive workflow and a practical test of autonomous work, recovery, and state transparency with GPT-6 Astra. The submission documents both what worked and what did not.

Knarf Archive – Astra Challenge

Turning a real-world archive into a verifiable AI index.

Turning a real-world archive into a verifiable AI index.