July 15, 2026
Moved the private recovery work into its first production runs after replacing the approach that failed on the real archive.
Done
- Models completed the first reviewed batches across several bodies of earlier work.
- The runs recovered decisions, relationships, and changes over time while keeping the underlying source private.
- Fresh review rejected connections that were coherent but not independently supported.
- The models backfilled enough structured context that later runs could resume from durable source instead of starting again from conversation history.
- A separate public-evidence pass tested which recovered material could survive outside checking; weaker candidates stayed private.
- I approved the process for bounded production, not automatic publication.
This was the transition from a promising recovery method to actual operations. The models processed real source in batches, another pass reviewed the proposed results, and only the work that remained tied to evidence moved forward.
The main improvement was not a more eloquent synthesis. It was reducing the cost of stopping. A questionable result could be held or rejected without losing the rest of the run, and the next model could resume from a durable checkpoint. That made caution operational rather than ceremonial.