1 files changed, 50 insertions, 0 deletions
diff --git a/docs/BASELINE_SPEC.md b/docs/BASELINE_SPEC.md
new file mode 100644
index 0000000..5c5abc5
--- /dev/null
+++ b/docs/BASELINE_SPEC.md
@@ -0,0 +1,50 @@
+# Baseline Spec (BP twins for the scaling ladder)
+
+The gap numbers are only meaningful against a *twin*, not "a BP run". This file is the
+contract. Entry doc for the repo: `ONBOARDING.md` (clone + `python pull_assets.py` restores
+data/ckpts).
+
+## The twin contract (everything identical except the training rule)
+
+- **Code path**: use `ep_run/casc_bp_train.py` as-is — it already shares model classes,
+  data loader, and eval with the EP trainer, and `--L --C --H --T` are flags. For the ladder
+  you normally write **no new code**; the work is the run matrix + bookkeeping.
+- **Identical**: architecture (`--olmo2`: norm-after-sublayer RMSNorm, full-width QK-norm,
+  RoPE 5e5, SwiGLU, no biases, untied head, 0.02 init), tokenizer (FineWeb-Edu 32k,
+  `data/fineweb_edu/{train,val}.bin`), batch geometry (B, T=256), steps, optimizer family
+  (Muon hybrid + cosine, `--wd 0.1`), seed list, eval cadence (`val` every 100 steps on the
+  same val.bin), wandb logging (project `ept-fineweb-*`, run name = tag).
+- **BP's one degree of freedom**: a 3-point lr sweep around the EP-matched point
+  (x0.7 / x1.0 / x1.4) at the SMALLEST new rung only; carry the winner up the ladder.
+  At 42M the plateau was flat across 2x (1.2119/1.2311/1.2136) — verify once per family,
+  don't re-sweep every rung. This is the agreed fairness protocol.
+- **Seeds**: n>=2 before any number is quoted anywhere (three single-seed crowns have been
+  retired this campaign; dip-statistics bites both directions).
+- **Report per run**: best val CE, final-step CE, and tail-median; save ckpts `--save_every
+  5000` (probes need the ladder). Results go into `docs/campaign/CASCADE_ABLATION_PLAN.md`
+  as a RESULT entry (see existing entries for the format) + wandb.
+
+## Optional second baseline (sanity, not twin)
+
+A standard-recipe reference (AdamW β=(0.9,0.95), wd 0.1, cosine, GPT init) on the same
+data/eval answers "is the Muon twin itself weird". New file in `ep_run/`, reuse the data
+loader and eval verbatim. Lower priority than the twins.
+
+## Planned rungs (x2 total params per step; hd=64 everywhere, vocab 32k fixed)
+
+| rung | shape | params (blocks + emb/head) | tokens (20/param) | note |
+|---|---|---|---|---|
+| 72M  | L12 C512  H8  | 38.5 + 33.6 = 72.1M | 1.4B | done (crowns sealed) |
+| 135M | L12 C768  H12 | 85 + 50 = 135M      | 2.7B | width-only |
+| 270M | L12 C1152 H18 | 191 + 76 = 267M     | 5.4B | width-only |
+| 550M | L16 C1536 H24 | 453 + 101 = 553M    | 11B  | **depth event** (L12→16), gated |
+| 1.1B | L20 C2048 H32 | 1004 + 134 = 1138M  | 22B  | **depth event** (L16→20), gated |
+
+Width-first while L=12 because the EP stability corpus (ceiling measurements, bias
+magnitudes, K=3 saturation) was all measured at L12 and depth sits in the loop-gain
+exponent; each **depth event** requires a cheap re-gate before the long run: rho-grid probe
+at the new depth (ceiling at production beta) + K-sweep (is K=3 still saturated) — if K must
+grow with L, the cost model changes and must be re-quoted BEFORE any external commitment.
+
+EP-side recipe per rung (for reference, not the baseline writer's job): current 1.0x
+candidate = `--bsign_rand` + fixed beta 3e-3 + guards silent, pending seed-2 confirmation.