blob: a2da91d90eb8201e1ce33375c56b362720e5bf29 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
|
# Phase 10A Memo: Prefit Threshold Curve
**Date**: 2026-03-26
**Config**: CIFAR-10, L=4, d=256, t0=5, blend_075, seed=42
## Question
How much offline prefit does Vec need before blend handoff helps?
## Answer: NONE. Even random Vec with blend(0.75) outperforms DFA by +1.3%.
| E_prefit | Gamma_frozen | rho_frozen | final acc | diff vs DFA |
|----------|-------------|-----------|-----------|-------------|
| 0 (random) | -0.005 | 0.014 | 0.324 | **+1.3%** |
| 15 | 0.002 | 0.011 | 0.323 | **+1.2%** |
| 60 | -0.001 | -0.009 | 0.325 | **+1.4%** |
**This is Case C**: very weak (or zero) prefit suffices for blend to beat DFA.
## Critical Reinterpretation
The Phase 9A success was NOT due to Vec learning useful credit. The frozen Gamma/rho are near zero at all prefit levels. The benefit comes from **blending DFA with any additional signal** — even random noise through a VectorCreditNet provides a regularization/diversification effect that improves over pure DFA.
This means:
1. The "cold-start paradox" narrative was partially wrong — Vec doesn't need to be good to help
2. The blend mechanism itself is the active ingredient, not Vec's credit quality
3. Phase 9A's +1.5% was not evidence that "Vec credit is useful online" — it was evidence that "blended updates regularize better than pure DFA"
## Implication
The next question is: is this just a regularization artifact (any noise helps), or does Vec's structure matter? This should be tested by comparing blend with random Vec vs blend with random noise of same magnitude.
|
