diff options
Diffstat (limited to 'docs')
| -rw-r--r-- | docs/campaign/FINDINGS.md | 23 |
1 files changed, 23 insertions, 0 deletions
diff --git a/docs/campaign/FINDINGS.md b/docs/campaign/FINDINGS.md index d3da6bb..4b78f78 100644 --- a/docs/campaign/FINDINGS.md +++ b/docs/campaign/FINDINGS.md @@ -737,3 +737,26 @@ docs/campaign/CASCADE_ABLATION_PLAN.md. Scripts: cascade_probe.py (schemes incl. casc_eq_train.py (v5: adaptive β via σ̂_tok, quality-governed K, cos telemetry). Negative results recorded: govfloor rr_floor 2/3 lethal at C512 (free phase must stay resreg-free); governor certification needs a ρ>0.95 lower bound (certified a ρ=0.80 corpse). + +## 2026-07-11 — K3 SEALED: matched-tuning PARITY (n=3), and the honest arc that got there + +**Final table (C256 L6 T256, TinyStories, 4k steps, both at tuned lr 1e-3, tok_init 0.02):** +- equilibrium fb-EP EXACT mode (dtop_every=1, fixed β 0.003): 2.0481 / 2.0687 / 2.0331 → **2.0500 ± 0.015** +- BP: 2.0503 / 2.0580 / 2.0506 → **2.0530 ± 0.004** +- ⟹ **statistically indistinguishable. "Standard transformer, no backprop, equilibrium-EP training + = BP quality at matched tuning."** cos(EP,BP) telemetry ≥0.9998 throughout all arms. + +**The three-act arc (all pre-registered readouts, no post-hoc stories):** +1. "EP beats BP by 0.074" (first C1) → KILLED by BP lr-sweep (BP@1e-3 → 2.0855; both had sat at + BP-suboptimal 3e-4). β-sweep showed NO β-growing dividend ⟹ prospective-configuration + regularization hypothesis not supported (β sweet zone 0.003–0.03; 0.1 breaks fb, 0.001 amplifies + /β noise). +2. "EP pays a ~4% tax" (3-seed matrix, v7 fast mode: 2.144 vs 2.053) → KILLED by the bisection arm: + the tax was ENTIRELY the v7 dedups (stale d_top at high lr). Exact mode restored parity. +3. Parity, n=3, sealed. Fast mode retained as a documented ~20%-compute/−4%-CE dial (exact = default). + +**Cost ledger (1080, C256 L6):** BP 8.8 it/s; EP exact ≈ 2.3–2.5 it/s (~3.6× BP); fast mode 2.68 it/s +(3.28×). Optimization headroom (compile on A-cards, ±-phase batching, bf16-relax pending A0.4) mapped. + +**Next: D-tier** — L12×C512 (~45M standard GPT-small) EP-vs-BP flagship + Muon arm (OmniOpt sweet +spot 60–350M; Muon = digital-side, free on GPU, EP-interaction unmeasured). |