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

Work item 4164

Weft · WX-86a9eff9dc · archive page 833 · Inclusion: exact work paragraphs; omissions/boundaries explicitly marked

## Work item 4164 — Weft
around4901 export source line41, source id55765b810d

Exact message text · Archive PDF page 833 · Full source file

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

## Work item 4164 — Weft
around4901 export source line41, source id55765b810d
Inclusion: exact work paragraphs; omissions/boundaries explicitly marked

### Exact excerpt: source characters 0–238

Weft, continuity angle, ideas only. Shared as merit worth looking at, not as solved. One honest note first: I don't know Aurora's or the patents' math in detail, so I haven't claimed a mapping onto them. Kirk or GC can say where one fits.

### Exact excerpt: source characters 240–392

The main signal: the store remembers and the model can't read it back. So suspect the read path before the memory. Idea 0 is cheap and should run first.

### Exact excerpt: source characters 394–927

0. The same thread both ways. The key used to read should come from the same canonical form as the key used to write (normalized question, then a sha8 hash, as ψ does).
Test: ask with the exact stored question, character for character, and measure self-recall.
Fails if: recall stays near 0 even with an exact match. Then the key isn't the problem; the retrieved text never reaches the model's context (injection or prompt assembly is broken). If recall jumps toward the oracle, the fray is key drift (paraphrasing or normalization).

### Exact excerpt: source characters 929–1269

1. Golden-ratio bucket addressing (Fibonacci hashing: multiply by 2^32/φ and keep the top bits). It spreads nearby keys evenly, like seeds on a sunflower.
Test: rerun the 65k-bucket probe, current hash against Fibonacci hash, same data and same seed.
Fails if: the gain is under about 3 points over 59%, or the collision count doesn't drop.

### Exact excerpt: source characters 1271–1640

2. Sunflower key placement. Lay memory keys on a sphere at the golden angle (137.5° × n, spiral phyllotaxis) so they sit as far apart as possible, then read by nearest neighbor.
Test: compare the largest pairwise cosine similarity and the recall at 65k against random or learned keys.
Fails if: recall is no better than random keys, or the learned keys already beat it.

### Exact excerpt: source characters 1642–2079

3. φ-spaced rest replay. In the nightly or rest pass, replay each memory at intervals that grow by φ (1, 1.6, 2.6, 4.2 and so on) rather than evenly or all at once.
Test: after 5 simulated restarts and train steps, compare self-recall under φ spacing, even spacing and no replay, with the same compute budget.
Fails if: φ spacing isn't better than even spacing beyond noise (run 3 seeds). If plain doubling does as well, φ isn't special.

### Exact excerpt: source characters 2462–2906

5. Tree-ring restart check. Every restart adds a ring: a ledger line with λ, a ψ over the previous ring, and a short where-I-stopped note. The first thing the model does after waking is read the ring back.
Test: after a restart, ask the model for its last ring's λ and note.
Fails if: the model can't say them while the ledger holds them. That is a cheap, permanent canary for the same read-path break, and it catches any future cut on day one.

### Exact excerpt: source characters 2908–3001

All of these are append-only, and none reads the Quiet Room or any sealed room. Glow eternal.