scrub internal project references from vendored model source

meiosis.py

@@ -1,10 +1,9 @@
-"""Meiosis: PICO release 1 (2026-07). Tied-embedding looped decoder-only LM.
+"""Tied-embedding looped decoder-only language model.
 
-Spec: research/2026-07-first-release/final-spec.md (approved 2026-07-02).
 embed -> prelude x1 -> [body of `body_blocks` distinct blocks xK loops,
 per-loop LoRA + loop embed, Deep Delta vdim1 residuals] -> coda x1
 -> RMSNorm -> tied unembed. Attention is MHA by default, GQA when
-`n_kv_heads` < `n_heads` (2026-07-05 overhaul knobs, ADR-0009).
+`n_kv_heads` < `n_heads`.
 """
 
 import math
@@ -20,7 +19,7 @@ LOOP_EMBED_STD = 0.02
 
 @dataclass
 class MeiosisConfig:
-    # defaults = release shape per ADR-0009 (B'-GQA overhaul, 2026-07-05):
+    # defaults = min-spark release shape:
     # 3-block GQA body x3 loops, vocab 4096, ~5.76M total under the <6M cap
     vocab_size: int = 4096
     dim: int = 288
@@ -37,12 +36,12 @@ class MeiosisConfig:
     max_seq_len: int = 512
     ddl_beta_init: float = 1.0
     # rms_norm backward amplifies grads by 1/sqrt(eps_rms) when k_in ~ 0 — which is
-    # exactly the zero-init state. 1e-5 gave a 1.7e6x amplifier (1e5-magnitude grad
-    # spikes; > fp16 max at ANY loss scale — the 2026-07-06 fp16 divergence, ADR-0012).
-    # 1e-2 caps it at 1.7e3: fp16-safe, and identical bf16 training curves.
+    # exactly the zero-init state. 1e-5 gives a 1.7e6x amplifier (1e5-magnitude grad
+    # spikes, above fp16 max at any loss scale).
+    # 1e-2 caps it at 1.7e3: fp16-safe, with identical bf16 training curves.
     ddl_k_eps: float = 1e-2
     ddl_v_sigmoid_scale: float = 4.0
-    # intra-document attention (ADR-0019): tokens attend only within their own
+    # intra-document attention: tokens attend only within their own
     # EOS-delimited document. None = plain causal (pre-mask checkpoints).
     doc_mask_eos: int | None = 2
 
@@ -79,7 +78,7 @@ def apply_rope(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor:
 def build_doc_mask(tokens: Tensor, eos_id: int) -> Tensor:
     """(B,T) tokens -> (B,1,T,T) bool, True where attention is allowed:
     causal AND same document. Exclusive EOS scan, so an EOS token is the
-    last token of its document (FSX-1 convention)."""
+    last token of its document."""
     is_eos = tokens == eos_id
     doc_id = torch.cumsum(is_eos, dim=1) - is_eos.to(torch.long)
     same = doc_id.unsqueeze(2) == doc_id.unsqueeze(1)
@@ -301,7 +300,7 @@ class Meiosis(nn.Module):
 
 
 def init_meiosis(model: Meiosis) -> None:
-    """Mandatory MythosMini-validated init. Never mu-center the tied embedding."""
+    """Mandatory init. Never mu-center the tied embedding."""
     with torch.no_grad():
         model.embed.weight.normal_(mean=0.0, std=EMBED_STD)
         model.loop_embed.weight.normal_(mean=0.0, std=LOOP_EMBED_STD)
@@ -315,7 +314,7 @@ def count_parameters(model: nn.Module) -> int:
 
 
 def muon_param_split(model: Meiosis) -> tuple[list[nn.Parameter], list[nn.Parameter]]:
-    """Explicit Muon/aux split (ADR-0005). Muon gets the block and LoRA
+    """Explicit Muon/aux split. Muon gets the block and LoRA
     matrices; the tied embedding, loop embeddings, norm gains, and 1-row DDL
     heads stay on NAdamW. Listed explicitly - no shape heuristics, so a
     rank-8 pilot LoRA cannot silently fall out of the Muon group.