# Why Assd/Carta H improved vs older models

Inventory date: 2026-11-09. Numbers from measured eval scripts only.

## Headline structure

- **`soft_hp50 × AccFuse↔linear_hp30 tip`** at **p=0.998 / sc=1.0**: LODO **0.8549 / 78.43**
  (operating milder **p=0.996 / sc=0.8**: LODO **0.8537 / 77.79**; AASD **0.6431**; Carta **0.6432**).
- Prior tip **AccFuse↔linear_hp30 α=0.50** @0.996/0.8: LODO **0.8011 / 71.43**; AASD **0.6421**; Carta **0.6522**.

## What H is

Official scorer: cohort **H = 2·(Acc/100)·L / ((Acc/100)+L)** — harmonic mean of sample Acc (fraction) and switch-latency score **L** (τ₀=60 s). HMM `p_stay` / emission `scale` selected on LODO only.

## Mechanism (why external H beats stem-family)

1. **Stem / IRM↔synth / SWA** clear synthetic LODO (H≈0.87–0.88, Acc≈82–83) but carry **slow signed bias** that synthetic schedules reward. On real AASD/Carta switches that bias hurts **L** and Acc → Carta H≈0.55–0.57, AASD H≈0.62.
2. **Causal τ=30 one-pole HP** on 4-seed logits (`linear_hp30`) strips that near-DC drift. Measured trade: LODO drops to **0.7899/69.21** while Carta rises to **0.6520/60.54** and AASD to **0.6422/53.12** (vs `linear_mean` Carta **0.6454**).
3. **AccFuse (SW2↔BMMS raw w=0.5) ↔ linear_hp30 α=0.50** restores LODO Acc (~+2–3 pt vs hp30 alone) while **keeping Carta H≥0.652**. vs stem raw fuse: Carta H **+0.10**.
4. **soft_hp50(stem) @ w=0.15** into that tip buys LODO **0.85+/78%** by importing lightly HP’d stem Acc; Carta/AASD fall to **~0.643** (Pareto wall ≈0.007 below H≥0.65 once LODO clears the new bars).

## Gap to joint gates

Need **LODO H>0.85 ∧ Acc>75** and **Assd H≥0.65 ∧ Carta H≥0.65**. Soft stem mixing that clears LODO tops out at external H≈0.644. Tip alone keeps Carta≥0.65 but LODO only ~0.80/71. Further soft-w / capped-resid / agree-gate / light-GRU mixes are closed. Need a **structurally different bias control** (preserve tip DC while importing stem AC, or train-time HP head, or emission-only gating).
