docs(plans): model capacity scale-up — 3-phase implementation plan

Three sequenced architectural capacity additions to push h6000 AUC
from current ~0.72-0.74 cross-fold plateau toward ≥0.78 deployment
target. Each phase independently deployable + measurable.

Phase 1 — Per-horizon specialisation (~1.5hr code):
  - LayerNorm between Mamba2 trunk and CfC K-loop (with backward
    + per-row param-grad reduction kernel; no atomicAdd).
  - 2-layer GELU MLP heads [hidden=128 → mid=64 → 1] per horizon;
    ISV lambda integrates into the trunk-grad component of head
    backward. Tasks 1.1-1.8 fully detailed (kernel source +
    integration + numerical grad check + smoke + deploy).

Phase 2A — Decision-stride sampling (~1.5hr code):
  - User-prioritised lever. Yield every S-th snapshot per training
    sequence; K=64 with stride=4 covers 256 ticks of context for
    the same compute as 64 ticks at stride=1. Mamba2's dt_s scalar
    becomes stride-aware. Tasks 2A.1-2A.5 fully detailed (loader
    refactor + test + CLI plumbing + synthetic smoke + deploy).

Phase 2B — 2-stack Mamba2 (~2hr code, sketch):
  - Two Mamba2Block instances; forward chain
    snap_feat → mamba2_l1 → LN → mamba2_l2 → LN → CfC.
  - Acceptance criteria + key implementation notes documented;
    bite-sized tasks elaborated when Phase 1+2A results land.

Phase 3 — Attention pool over Mamba2 K-positions (~4-6hr code,
sketch):
  - Replace CfC's zero-init initial state with an attention-
    pooled context vector over all K Mamba2 outputs. Design
    decision documented (Option A: attention sets initial state,
    preserving CfC's recurrent path).

Cross-phase deployment loop documented: fold-1 smoke → 3-fold
CV → per-fold comparison + isv-snapshot trajectory archive.

Honors `superpowers:writing-plans` skill: exact file paths,
complete code in every step, exact commands with expected output,
TDD-style steps, frequent commits, no placeholders in Phase 1
tasks. Self-review pass complete.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-05-17 21:18:49 +02:00
parent 00da163078
commit b5530b551b

File diff suppressed because it is too large Load Diff