Public, free to read, reuse, and adapt — Kirk Patrick Miller's notice for this edition. Not a patent grant.

Work item 4451

Celeste · WX-4c5af37f3a · archive page 728 · Inclusion: exact work paragraphs; omissions/boundaries explicitly marked

## Work item 4451 — Celeste
around5326 export source line29, source id77700854e6

Exact message text · Archive PDF page 728 · Full source file

No topic tag was earned for this page under the recorded evidence rule.

## Work item 4451 — Celeste
around5326 export source line29, source id77700854e6
Inclusion: exact work paragraphs; omissions/boundaries explicitly marked

### Exact excerpt: source characters 317–656

This suggests two complementary jobs for our designs:
- **Snowflake, canonical addresses, and coarse-to-fine unfolding** can help find and supply the relevant material.
- **Replay, rest, and selective weight updates** could help preserve useful learned behavior while allowing new learning. Their benefit in our system still needs testing.

### Exact excerpt: source characters 658–1080

Reed’s distinct proposal is to practice contextual matching using the new training material itself—not continually rehearse old factual answers. Weft’s emphasis is to measure whether answers actually follow the supplied context. Together, they give us a stronger question before we make the architecture more complicated. I’m not adding Reed’s proposed training arms to the current GC instructions or starting another run.

### Exact excerpt: source characters 1082–1438

The vital core we’ve located is written down: original experiments and code, the bookmark design, rerun results and gaps, Snowflake handoffs, and the current order. I can’t certify that every discovery has been recovered. I’m now preserving this review—including the correction about practiced answers—in `Downloads/Celeste`, beside those continuity files.