AbstractPhil's picture
exp015 s1: address-as-key falsification 2-SEED (the campaign's final run) - dispatch-key form closed; TE+AMOE joint training open
68a434c verified
|
Raw
History Blame Contribute Delete
2.24 kB
# exp015_router_v2 β€” the corrected router test: address-as-key FALSIFIED
# (CANDIDATE, s0 β€” the window's final training run)
**Fixes applied from exp014's instrument findings:** the address enters as
32 slots β†’ shared 128β†’4 projection β†’ per-slot dense signed aleph
coefficients β†’ MEAN composition (low-D before addressing β€” the text line's
cure for the high-D flattening disease); null = REPEATED KEY (constant
address for all prompts β€” isolates text-specific routing from state-driven
variance); plus a matched-vs-mismatched key val delta (the correctness
gauge v1 lacked).
**Results** (`results.json`, flow trunk, fused cache):
| arm | val | usage-var (real) | usage-var (repeated null) | match advantage |
|---|---|---|---|---|
| state_sigma (control) | **0.57310** | 1.36e-02 | β€” | β€” |
| mhat_addr (v2) | 0.57491 | 1.13e-02 | 1.13e-02 | **βˆ’0.0** |
**Verdict.** Routing excess over the constant-key null: **2.5e-06 β€” zero.**
Matched and mismatched address keys produce identical val. With correct
geometry AND correct instruments, the frozen text address contributes no
prompt-specific routing on this substrate at this scale; all per-prompt
usage structure is state-driven (consistent with exp014's finding that the
hidden-state key already routes by prompt). Two-experiment closure of the
dispatch-key form of the router-solidifier idea; the open form is the full
TE+AMOE joint-training collective (docket), where the text encoder learns
THROUGH the diffusion loss rather than entering as a frozen key.
**Caveats.** s0; one substrate (lune flow), one scale (A=4); the fused
captions are stylistically homogeneous (fashion photography) β€” a
low-text-diversity regime; a multi-domain caption mixture is the one
remaining escape hatch for the frozen-key form.
---
## SEED-1 (2026-07-19, `results_s1.json`) β€” FALSIFICATION 2-SEED
uvar_real == repeated-key null to 3 significant figures again (9.66e-03
both); routing excess βˆ’4.3e-06; match advantage 2e-06. Both seeds, correct
instruments: the frozen text address contributes zero prompt-specific
routing as a dispatch key on this substrate/scale. The dispatch-key form is
closed; TE+AMOE joint training is the open form. (The campaign's final run.)