exp006_sd15core_relay/README.md

exp006_sd15core_relay — relays on the CORE EPRED SD1.5 (CANDIDATE, s0)

Question. Does the relay recipe certify on the primary substrate — stock stable-diffusion-v1-5, epsilon objective on the shipped schedule, natural 77-token plain-English conditioning — and does it beat matched LoRA there?

Design. exp001's design transplanted to the natural paradigm: frozen fp32 stock UNet; three arms on one shared cache (4096/256 rows, latents + nl77 conditioning); epsilon-prediction loss with the stock scaled-linear schedule, t~U{0..999}, 10% CFG dropout; 3000 steps, batch 16, pure Adam wd=0, gradient checkpointing; paired val on FIXED per-row (noise, t), fp32. Round-trip judge via the standard StableDiffusionPipeline (stock scheduler) on exp000's rows, read against exp000b's sd15_epred_core baseline row (+0.1809).

Results (results.json):

arm	trainable	val eps-MSE	Δ vs frozen	round-trip gap
frozen	0	0.124788	—	(+0.1809 baseline)
relay_all16	3,227,568	0.121698	−2.5%	+0.2172
lora_r32	3,182,592	0.122674	−1.7%	+0.1596

Findings (candidate, s0).

1. Relay ≥ matched LoRA, 2-for-2 across substrates (here and exp001 on the Lune flow trunk). On the core, LoRA does help val (unlike Lune where it landed below frozen) — but stays behind the relay.
2. Grounding: relay gains, LoRA trades away. The relay arm improved natural-prompt round-trip grounding over the zero-shot core baseline (+0.1809 → +0.2172) while LoRA degraded it (+0.1596). Same direction as exp001; stronger effect.
3. Gate dynamics are substrate-dependent: core-epred relay gates GREW (0.0474 init → mean 0.0700, max 0.0986 — the "trunk opts in" mechanism from the text line), where Lune relay gates shrank (exp001, mean 0.0432). Neither sits in the LM-line 0.012–0.03 band.
4. Toggle law held post-train, bit-exact (toggled-off val ≡ frozen).

Caveats. Single seed; 3k-step short-train (machinery certification, not a capability claim); one LoRA placement/rank; round-trip n=24. Cost ≈ 3.1 GPU-h. Checkpoints: relay_all16_s0.pt, LoRA in the run dir (ships on request — the relay stack is the line's artifact).

---

SEED-1 REPLICATION (2026-07-17, results_s1.json) — PROMOTED TO 2-SEED CONFIRMED

Same cache/val; training stochasticity reseeded. relay 0.121709 (s0 0.121698), lora 0.122673 (s0 0.122674), frozen identical. Relay beats frozen AND matched LoRA in BOTH seeds; round-trip relay +0.2206 (s0 +0.2172) vs lora +0.1889 (s0 +0.1596); gate growth replicates (mean 0.0707 vs 0.0700). No seed lottery. Scope: 2 training seeds on one shared cache; data reseeding untested.