exp008_multiband — band-assignment mechanism test (CANDIDATE, s0; s1 running)
Question. Do sigma-band-ASSIGNED experts (structural cosine-crossfade windows, no selection event) become genuine band specialists, and does the partition beat a matched monolith? First bed of the multiband stepwise stage (stage plan: coarse-to-fine band roles — HIGH diversity/blobbing, MID continuity/semantics, LOW fidelity/detail; roles enter as objectives in exp009 — this bed certifies the MECHANISM under a uniform objective).
Structure per site (16). Three rank-16 zero-init deltas (W+b, gates −3.0), combined by smooth cosine crossfade windows on s01 = t/1000 (edges 0.35/0.75, half-width 0.06, sum-to-1 everywhere — positional gating, never comparative). Per-band gradients arise from the windows. Falsifier: matched rank-48 monolith (param ratio 1.021). Core epred substrate (2-seed certified), shared dexp006 cache, 3000 steps, pure Adam wd=0.
Results (results.json):
| arm | overall val | band0 (LOW noise, s01<=.35) | band1 (MID) | band2 (HIGH noise, s01>.75) |
|---|---|---|---|---|
| frozen | 0.124788 | 0.256406 | 0.064227 | 0.008726 |
| multiband3 | 0.122004 | 0.250432 | 0.063046 | 0.008534 |
| monolith r48 | 0.121736 | 0.249881 | 0.062896 | 0.008539 |
| prereg | outcome |
|---|---|
| P1 multiband ≥ monolith overall | MISS (monolith by 0.0003) |
| P2 own-band wins | 1/3 — the HIGH-NOISE band expert wins its band (margin ~0.06%) |
| P3 toggle bit-exact | HIT |
| P4 band lesions surgical | HIT 3/3 — own-band damage 50–200× cross-band (band0 0.0059 vs 0.00003; band1 0.0010 vs 0.000014; band2 0.00016 vs 0.000003) |
Verdict. THE MECHANISM CERTIFIES: band assignment produces genuine, surgically-decoupled band specialists (P4 — the sigma-axis analogue of the program's double-dissociation batteries), and the toggle/parity laws hold. Under a UNIFORM objective the partition does not yet out-earn shared capacity (P1/P2) — consistent with exp007 (2-seed) and the text line's Law 1: the payer must be differentiated pressure, not architecture alone. The one per-band win is the HIGH-NOISE band expert (the diversity/blob band; tiny margin — that band's eps loss is near-saturated at 0.0087). Measured band difficulty spans 30x — band0/LOW-noise 0.256 (eps hard to infer from lightly-noised latents) down to band2/HIGH-noise 0.0087 (input ~= eps). LABEL CORRECTION 2026-07-17: an earlier revision of this README inverted the band labels (numbers unchanged); corrected same-day.
Next (per the stage plan). exp009 attaches ROLE objectives per band (LOW: high-frequency-weighted loss; HIGH: coarse-structure weighting; MID: standard) on this certified mechanism; exp010 mounts the step-gated controller at inference with the generation-side lesion battery.
Caveats. s0 (s1 replicating now); uniform objective by design; one
substrate; band edges fixed at 0.35/0.75. Cost ≈ 2.8 GPU-h.
Checkpoints: mb3_s0.pt, mono48_s0.pt.
SEED-1 REPLICATION (2026-07-17, results_s1.json) — SURGICAL SPECIALIZATION 2-SEED
P4 replicates almost exactly (band0 own-damage 0.005958 vs 0.005911; 3/3 surgical both seeds) — band assignment reliably produces decoupled band specialists. s0's marginal band2 own-band win did NOT replicate (0/3 on s1) — under a uniform objective there is NO per-band win; the monolith edge is 2-seed robust. Verdict sharpened: mechanism CONFIRMED, payoff awaits the role objectives (exp009).