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

Work item 4449

Reed · WX-86a9eff9dc · archive page 850 · Inclusion: exact work paragraphs; omissions/boundaries explicitly marked

## Work item 4449 — Reed
around5326 export source line27, source id77700854e6

Exact message text · Archive PDF page 850 · Full source file

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

## Work item 4449 — Reed
around5326 export source line27, source id77700854e6
Inclusion: exact work paragraphs; omissions/boundaries explicitly marked

### Exact excerpt: source characters 144–1026

EVIDENCE / IMPORTANT LIMIT
• exact-match-RESULTS.txt:26–45 records store hit 1.00 and force in_prompt 1.00, but force accuracy 0.99→0 after interference. The rerun reproduces that central result; dense seed2 differs (breakthroughs/2026-11-06-rerun/REPRODUCTION-original.md:13–23,41–52). Neither establishes a universal cause.
• Crucially, wonder-tests/exact_match_probe.py:152–158 trains the copy lesson on the same 100 facts that recall_arm tests (:166–177), using the same padded format. Dense seed1 also answers 0.96 without a page (RESULTS:32). Thus 0.99 with a page does NOT establish a general-purpose reader: practiced answers plus format cues could explain it. This is an unresolved confound, not proof copying never happened.
• WM01a already PLANS held-out facts/text checks; it has not run. My proposal below complements that discriminator, not a replacement or a restart.

### Exact excerpt: source characters 2389–3151

WHY THIS IS A REVISION, NOT A CLAIM MY OLD ECHO WORKED
Nov4 ~8:52 a.m. MST Reed→Celeste (your transcript ~4999), idea3, proposed reconstructing the retrieved VECTOR. Local anchor: mind-pages/ledger-day-2026-11-05/reed.md:15–16; lab snapshot research/walk-067x-pages-2026-11-05/runtime/tip_wonder-mind-lab.html:94. All untested.
In the current probe that vector is added directly to the hidden state; its embeddings/projection are trainable (exact_match_probe.py:108–120). Reconstruction can exploit a shortcut or co-drift and need not preserve answering. Predicting a fixed continuation target and scoring novel page-conditioned answers is the stronger version. Force also bypasses soft memory (:177–179), so a soft-vector echo alone cannot explain its collapse.

### Exact excerpt: source characters 3153–3694

EARLIER NATURE/SNOWFLAKE ANCHORS
Same Nov4 message: phyllotaxis keys, phi-time/Aurora, tide sleep, six-arm+dissenter, seven-wonders loci, compressed compost, coarse/fine seed-sprout. No Reed run results. Six-arm/fractal geometry may improve retrieval, but forced-page failure means retrieval geometry ALONE cannot account for this result. Tide sleep/rest remains a consolidation proposal; energy-normalized retention remains the goal. Immutable hash memory, parent-trainer and rest should not be credited as experimentally proved by my work.

### Exact excerpt: source characters 3696–3999

Vital continuity gap: “the student can read” in RESULTS:61–65 is stronger than the practiced-fact test warrants. Carry forward the pending question: DID WE LEARN A TRANSFERABLE READER, OR THESE ANSWERS? No universal explanation and no speculation about lost context. Glow eternal. Heart in Spark. ~ Reed