diff options
| -rw-r--r-- | ONBOARDING.md | 12 | ||||
| -rw-r--r-- | docs/campaign/CASCADE_ABLATION_PLAN.md | 9 |
2 files changed, 21 insertions, 0 deletions
diff --git a/ONBOARDING.md b/ONBOARDING.md index ce6e549..5cb8962 100644 --- a/ONBOARDING.md +++ b/ONBOARDING.md @@ -114,3 +114,15 @@ private repo. (The full `runs/` history is larger; `pull_assets.py` fetches the *Repo layout:* `ep_run/` = code (start at `lt_ep_train.py`) · `docs/` = the above · `scripts/` = standalone analysis/probe scripts · `assets/` = PDFs/figures · `refs/` = external paper texts · `archive/` = stale snapshots. + +## 2026-07-11 — THE CASCADE PIVOT (read this before the older sections) +The project's product architecture changed on 2026-07-09: we now train **standard multi-layer +transformers** (plain-forward inference, identical to any GPT) with **equilibrium-mode cascade-EP** +— layered energy E = Σ½‖z_l − f_l(z_{l−1})‖², free phase = one forward pass (exact), nudged phase = +3 forward-backward message-passing rounds to the nudged equilibrium, per-layer local θ-reads. No +backprop anywhere in training (formally audited; invariant test `ep_run/test_bp_free.py`). +**Sealed result: matched-tuning parity with backprop** (2.0500±0.015 vs 2.0530±0.004, n=3, C256 L6) +at ~3.6× GPU cost. Entry points: `ep_run/casc_eq_train.py` (trainer), `ep_run/cascade_probe.py` +(gradient gate), `docs/campaign/CASCADE_ABLATION_PLAN.md` (program), FINDINGS.md 2026-07-09→11. +The looped/weight-tied single block (everything below) remains the dynamics-physics testbed and the +subject of the aep-dynamics paper; the cascade line is the scaling/product path. diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md index 03e3b52..b6d8927 100644 --- a/docs/campaign/CASCADE_ABLATION_PLAN.md +++ b/docs/campaign/CASCADE_ABLATION_PLAN.md @@ -26,6 +26,15 @@ The looped-EP precedent multiplier was ~230× BP; the K-frontier decides whether --- +## STATUS 2026-07-11: K1+K2+K3 SEALED; D-tier in flight +- K1 exactness: cos 0.9998-1.0000 on-trajectory + BP-free formally audited (test_bp_free.py in repo). +- K2 cost: exact mode ~3.6x BP (v7); Sol audit says remaining eager headroom 5-10% (v8 queued). +- K3 quality: **matched-tuning PARITY n=3** (EP-exact 2.0500±0.015 vs BP 2.0530±0.004 @ C256 L6, + lr 1e-3 both). Arc: fake-win (lr artifact) -> fake-tax (v7 dedups) -> parity. Fast mode = documented + -4%CE/+20%speed dial. A0.4: TF32 free, bf16 production-only (cos 0.9427). +- D1a (L12xC512 45M): BP 1.917 sealed; EP + Muon arms landing. D1b long-run demo gated on D1a parity. +- E-tier: next in queue (softmax pathology / error-channel SNR / write pricing) -> Demo-0 spec sheet. + ## STATUS 2026-07-09 (same day): Tier 0 CLOSED GREEN via the zil scheme; C1 running - **Naive relaxation FAILS at depth** (the B1-lite sweep): jacobi K=40·L → cos 0.82 (L6) / 0.67 (L12) |