File size: 2,236 Bytes
4e7b91c 3074208 | 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 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 | # exp012_blob_supervised — blob supervision v1 + THE ROLE-ALIGNED GAUGE
# DISCOVERY (CANDIDATE s0; s1 running) + ANIMA R0b CERTIFICATION
**Blob supervision v1** (foreground-weighted LP-x0 on the HIGH band, λ=0.5):
HONEST NEGATIVE on both instruments — training-side blob gauge moved +0.03%
(0.111495 vs uniform 0.111528; P1 technical hit, economically nil); the
image battery (P4) showed mild dilution (grounding +0.1931 vs +0.2135;
HIGH-lesion LP-dominance 0.0662 vs 0.0727 — not sharpened). The eps term
dominates at this dose; stronger coupling or pure-band phases are the
follow-ups.
**THE KEEPER — the role-aligned gauge (in-bed reference gauging):**
| arm | HIGH-band foreground-LP-x0 gauge |
|---|---|
| frozen | 0.15549 |
| monolith r48 | 0.12339 |
| **uniform multiband** | **0.11153** |
| **blob multiband** | **0.11150** |
**The first gauge where multiband CLEARLY beats the matched monolith (~10%)**
— band specialists are better at foreground coarse structure in the
HIGH-noise band, exactly where the coarse-to-fine role map says they should
be. Every aggregate-eps comparison had hidden this. Band assignment's payer
is role-aligned MEASUREMENT — the specialists were already earning.
**ANIMA R0b (bundled: `r0b_probe.py`): ALL GATES GREEN.** 28/28 DiT blocks
carry relays (14,275,716 bf16 params), freeze-by-lr-0 exact, zero-init +
toggle exact. The gate CAUGHT a dtype-law violation live (fresh relays
attached fp32 on the bf16 trunk via a meta-tensor sniff; fixed to the
DECLARED dtype, fork commit on feat/aleph-adapter). exp004 training is
ready.
**Caveats.** s0 (s1 running); one λ; one coupling form. Cost ≈ 1.6 GPU-h
+ 0.2h battery + R0b probes.
---
## SEED-1 (2026-07-18, `results_s1.json`) — BLOB-ON-EPS 2-SEED INERT;
## THE GAUGE DISCOVERY HOLDS
s1: blob 0.112651 vs uniform 0.111528 (P1 MISS; s0 was +0.03%) — blob
supervision on the EPS substrate is inert-to-slightly-negative across
seeds, consistent with the conditioning hypothesis (x0 recovery divides by
vanishing sqrt(abar) at high noise). The multiband-vs-monolith role-aligned
advantage HOLDS both seeds (~0.112 vs 0.1234, ~10%). exp013 tests the same
coupling on the FLOW substrate (exact linear x0 recovery) — running.
|