# Lead optimizer session 2026-11-09

## Gates

| Gate | Target | Best this session | Status |
|---|---|---|---|
| LODO H | > 0.85 | **0.8549** (`soft_hp50` @0.998/1.0) | PASS (baseline) |
| LODO Acc | > 75% | **78.43%** | PASS (baseline) |
| Assd H | ≥ 0.65 | **0.6431** (soft_hp50 @0.996/0.8) | FAIL |
| Carta H | ≥ 0.65 | **0.6432** (soft_hp50 @0.996/0.8) | FAIL |

**Gates met: NO**

## Best retained structure

- Config: AccFuse (SW2↔BMMS w=0.5) ↔ `linear_hp30` α=0.50, × `soft_hp50(IRM↔synth w=0.5)` w_stem=0.15
- Code: `soft_hp50_accfuse_final.py`
- Results: `multi_agent_opt/results_soft_hp50_accfuse_final.md`
- Why Assd/Carta H beats older stem models: `multi_agent_opt/results_why_assd_carta_h.md`

## Why Assd/Carta H improved vs old models (brief)

Stem/SWA/IRM↔synth clear synthetic LODO (~0.87–0.88) but carry slow signed bias → Carta H~0.55. Causal τ=30 HP strips that bias (`linear_hp30` Carta **0.652**). AccFuse↔hp30 α=0.5 keeps Carta≥0.65 while lifting LODO Acc; soft_hp50@0.15 then buys LODO **0.85+/78%** at the cost of Carta/AASD → **~0.643** (Pareto wall ≈0.007 below 0.65).

## Tried this session (all FAIL joint; appended to ALREADY_TRIED)

1. AC-transplant tip+α(stem_hp−tip_hp) + mild rescore
2. Emission-σ gate on tip / soft_hp50
3. Switch-amplify dyn-pstay proxy
4. Light causal TF residual with hard HP (`models/light_hp_resid_tf` **deleted**)
5. ELM HP residual
6. AC-transplant of HP(soft_hp50)
7. Decision-fuse abandoned (no n128 on LODO)
8. Evidence |Δ|-gate tip↔soft_hp50

## Deleted failing artifacts

- `models/light_hp_resid_tf/` (entire dir after LODO 0.876 / Carta 0.608 fail)
