July 16, 2026
Ran the recovery work at substantially larger scale, fixed gaps found in review, and narrowed two separate projects before either could overclaim.
Done
- Models processed a much broader span of private work and produced reviewed records that connected present claims back to retained source.
- The work expanded beyond isolated examples into longer arcs: how projects changed, how decisions formed, and what later evidence did or did not confirm.
- Review found missing context and mistaken joins in several otherwise plausible records. The models repaired the source work and reran the affected reviews.
- The private fellowship application was checked against a fixed evidence checkpoint so later drafting could not quietly improve the underlying record.
- Models developed a pilot for documenting unwanted calls; adversarial review showed that the first version tried to prove and serve too many things at once.
- I narrowed that pilot around a smaller buyer and a clearer stopping rule.
The recovery system moved from examples to throughput. Models could work across a larger private corpus, preserve where each assertion came from, and send uncertain material back for repair without restarting the whole operation.
Scale made a familiar model failure easier to see: the more relevant fragments a model finds, the easier it is to assemble a story that feels complete before the evidence is. The review layer found several of those joins. Some were repaired; others stayed out.
The product pilot improved for the same reason. Its first version accumulated purposes faster than it accumulated proof. Cutting it down produced a smaller project with a better chance of answering a real question.