exp013_blob_flow/README.md

exp013_blob_flow — THE CONDITIONING HYPOTHESIS CONFIRMED (CANDIDATE s0; s1 running)

Hypothesis (Phil, 2026-07-18). "Blob likely will be less effective for eps than it will be for flow matching." Mechanism: eps-prediction recovers structure through x0̂ = (x_t − √(1−ᾱ)·ε̂)/√ᾱ — ill-conditioned at high noise (vanishing √ᾱ) exactly where blob supervision applies; rectified flow recovers x0̂ = x_t − σ·v̂ EXACTLY and linearly at every sigma.

Design. Identical coupling to exp012 (foreground-weighted LP-x0, HIGH band, λ=0.5), identical fused data + blob targets, identical multiband beds — only the substrate changes: base_lune (flow objective, its judged-best nl77 conditioning). Arms: frozen | uniform-mb3 | blob-mb3; in-bed gauges on all arms.

Results (results.json):

arm	common flow-MSE	HIGH-band blob gauge
frozen	0.62989	0.18447
uniform mb3	0.57249 (−9.1%)	0.12987
blob mb3	0.57379	0.12220

prereg	outcome
P1: blob-gauge move ≥5× the eps reference (+0.03%)	HIT — −5.9% relative, ~200×
P2 common gauge undegraded (<0.5%)	HIT (+0.23%)
P3 toggle bit-exact	HIT

Verdict (candidate, s0). The SAME supervision that is 2-seed inert on the eps substrate (exp012) bites hard on the flow substrate — the conditioning of the x0 recovery is the controlling variable, exactly as hypothesized. Consequences: (1) blob/structural supervision belongs on FLOW-objective trunks (or v-pred); (2) exp012 is reframed as this pair's conditioning control; (3) the undertrained lune trunk pays adapters −9.1% (largest adapter effect measured in the line — headroom scales inversely with trunk training, consistent with the "undertrained exemplar" framing).

Caveats. s0 (s1 running); one λ; lune ≠ core substrate (the flow-vs-eps contrast rides on different trunks — a v-pred core arm is the clean follow-up); blob gauge scale differs across objectives (relative moves are the comparison, stated as such).

---

SEED-1 (2026-07-18, results_s1.json) — PROMOTED: 2-SEED CONFIRMED

s1 relative blob-gauge move −3.7% (s0 −5.9%) — both seeds pass the preregistered 5× bar by ~125–200× the eps reference; common gauge clean both seeds. THE CONDITIONING LAW (2-seed): structural/blob supervision pays where x0 recovery is linear (flow: x0 = x_t − σv) and is inert where it divides by vanishing √ᾱ (eps at high noise). Scope: lune flow trunk vs core eps trunk (different trunks — v-pred same-trunk arm remains the sharpest follow-up); 2 training seeds, shared cache.

---

DOSE-RESPONSE POINT (2026-07-19, λ=2.0, results_lambda2.json)

Monotone: λ=0.5 → −5.9% blob gauge at ~0% common cost; λ=2.0 → −8.4% at +0.9% common cost (now past the 0.5% P2 bound). The coupling-strength trade surface is real and mapped at two points; λ≈1 interpolation and pure-band-phase training are the open refinements. (Ops note: this arm reused the s0 results filename on-pod — the shipped s0 file is canonical; the config's lambda field disambiguates.)

---

THE DOSE CURVE COMPLETE (2026-07-19, results_lambda1.json)

λ	blob-gauge move	common-gauge cost	within 0.5% bound
0.5	−5.9%	~+0.2%	yes
1.0	−8.3%	+0.4%	yes — the operating point
2.0	−8.4%	+0.9%	no
Monotone with saturation between λ=1 and 2: λ≈1 captures essentially the
full blob-gauge gain inside the preregistered common-gauge tolerance. All
single-seed dose points on the 2-seed-confirmed conditioning law's flow
substrate.