diff options
| -rw-r--r-- | docs/campaign/FINDINGS.md | 10 |
1 files changed, 8 insertions, 2 deletions
diff --git a/docs/campaign/FINDINGS.md b/docs/campaign/FINDINGS.md index 6ab1339..d3da6bb 100644 --- a/docs/campaign/FINDINGS.md +++ b/docs/campaign/FINDINGS.md @@ -721,8 +721,14 @@ nudge +β·CE at the top, relax states to the nudged equilibrium, ∇θ = local **Result (diagC, tied + 0.02 init, equilibrium fb-EP, NO regularizers, NO guards triggered): best val CE 2.9021 in 1500 steps, cos(EP,BP) pinned at 1.0000 throughout, K=3 never raised** — -already below the bad-init BP twin's 4000-step 2.9746. Final same-init C1 pair (casc_bp6i vs -casc_eqv5, 4k steps) running. +already below the bad-init BP twin's 4000-step 2.9746. + +**FINAL C1 VERDICT (same-init pair, 4k steps, seed 0): equilibrium fb-EP 2.2949 vs BP 2.3694 — +EP BEATS BP by 0.074** (cos pinned 1.0000 the whole run, zero skips, K=3 never raised, wall-clock +3.9× BP: 2.23 vs 8.79 it/s on the same 1080). K1 (exactness-on-trajectory), K2 (GPU cost), and K3 +(training parity, exceeded) are all bought. Same EP>BP flavor as the looped-EP 1.7065-vs-1.7921 +flagship — second independent instance of the pattern. Remaining: K4 depth rungs (L12 C512, D-tier) +and K5 analog tolerances (E-tier). **Costs & implications:** equilibrium-mode ≈ 4× BP (fb K3 single-sided) ⟹ $20k AWS ≈ 3–4B-param Chinchilla run in TRUE-EP mode. Optimization headroom mapped (±phase batching, red-black grouped |