AbstractPhil's picture
exp012 s1: blob-on-eps 2-seed inert (conditioning hypothesis consistent); role-aligned gauge advantage holds both seeds; exp013 flow test running
3074208 verified
|
Raw
History Blame Contribute Delete
2.24 kB

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.