# results_seed_disagree

**Claim:** `seed-disagree-gate` — pairwise **4-seed** causal disagreement soft/hard gate on fused evidence (raw mean / `linear_hp30` / median). **NOT** fuse↔hp30 stream assignment.

**Honesty:** At each frame `disagree = mean_{i<j}|s_i−s_j|` on raw or causal-hp30'd seeds; optional EWMA τ∈{0,1,4,8}s @ 0.25 Hz hop. Soft: `e'=e·sigmoid(α(thr−d))`. Hard: if EWMA d > thr → hold last accepted or emit 0. Thr = train CORPUS p50/p75 of disagree (no AASD/Carta labels). Select mean LODO **Acc then H** @ HMM **0.994/0.8**.

**Compute:** CPU-only (~20 min). No `gpu.lock`. Artifacts: `seed_disagree_gate.py`, `results/research/seed_disagree_gate.json`, `models/seed_disagree_gate/selection.json`, `experiments/seed_disagree_gate.log`.

## LODO (selection)

| Variant | mean H / Acc |
|---------|--------------|
| **hard_hold_raw_mean_t8_p75** (selected) | **0.7999 / 75.02** |
| hard_zero_raw_mean_t8_p75 | 0.7946 / 74.84 |
| raw_mean (ungated) | 0.7932 / 72.68 |
| med4 | 0.7933 / 72.63 |
| hard_hold_hp30_t8_p75 | 0.7906 / 71.18 |
| linear_hp30 (ungated) | 0.7797 / 68.15 |
| soft_hp30_* (best Acc) | ≤0.7811 / 68.32 |

Hard hold on **raw_mean** lifts LODO Acc ~+2.3 vs ungated raw; soft gates on hp30 are ~flat vs `linear_hp30`.

## External (headline = selected)

| Variant | AASD H / Acc | ΔAcc vs 53.12 | Carta H / Acc |
|---------|--------------|---------------|---------------|
| **hard_hold_raw_mean_t8_p75** (selected) | **0.6198 / 51.12** | **−2.00** | **0.5896 / 54.63** |
| raw_mean | 0.6417 / 52.91 | −0.21 | 0.6454 / 59.28 |
| linear_hp30 | 0.6422 / 53.12 | 0.00 | 0.6520 / 60.54 |
| soft_hp30 (typical) | ~0.64 / ~53.0 | ~−0.1 | ~0.65 / ~60.5 |
| hard_*_hp30 | ≤52.30 Acc | ≤−0.8 | regress Carta |

Best AASD Acc in the whole grid remains ungated **`linear_hp30` (53.12)**. Soft shrink barely moves external; hard hold that wins LODO **hurts** AASD/Carta.

## Verdict

- Honest LODO selection → `hard_hold_raw_mean_t8_p75` (LODO **0.7999/75.02**, AVGC **0.7212/68.89**).
- AASD Acc **51.12** vs portfolio external best **53.12** (Δ **−2.00**). Carta regresses to 54.63.
- Seed-disagreement gating improves train Acc on raw mean but does **not** close AASD Acc→60; soft hp30 gates ≈ identity.
- Open targets unchanged: AASD Acc>60 still −6.88 under `linear_hp30`; AVGC still short of 0.85/78.

## Do not repeat

Do not rerun this pairwise-seed soft/hard gate grid (raw|hp30|median × τ×α×p50/p75 thr × hold/zero) at HMM 0.994/0.8. Distinct from exhausted fuse↔hp30 MoE/emis-LL/MMD/xstream families — still do not repeat those either.

## NEXT IDEA

**Causal seed-reliability weights** (not a gate-to-0): past-only inverse-variance / EWMA agreement weights over the 4 seeds → continuous weighted mean (always emit; never hard hold/zero). Orthogonal to disagreement *attenuation*. Alternate: wait for **ms-feat-coral** (higher-dim) if still in flight — 1-D logit post-processing is saturated.
