# results_evidential_tip_stack

**Claim:** `evidential-tip-stack` — train-only **linear / logistic / Dirichlet-evidential stacking** over frozen tip stems (**focal + synth + IRM**). Leave-AVGC fits on non-AVGC train only.

**Distinct from** closed `multi-stem-attn-pool` (learned EWMA→MLP soft-attn gates), `causal-ds-temp` / `isotonic-ds-fuse` (PP on tip fuse), agree-gate / conf-resid / fixed-w IRM↔stem.

**Honesty:** Train labels only. Never AASD/Carta for selection. Select mean LODO **H then Acc** @ **0.994/0.8**.

**Compute:** CPU; reuse focal + fuse_conf_resid IRM/synth caches. Code: `evidential_tip_stack.py`. JSON: `results/research/evidential_tip_stack.json`.

## Selection

| Rule | Pick |
|------|------|
| mean LODO H then Acc (stack variants) | **`dirichlet_evid`** |
| Tip to beat (held_T_nll_med) | **0.8245 / 75.64** |

Ranked H: dirichlet_evid → nnls_simplex → logistic_stack

## LODO (preset)

| Variant | mean H / Acc | AVGC H / Acc | fit@AVGC | selectable |
|---------|--------------|--------------|----------|------------|
| focal_synth_w05 | 0.8821 / 83.36 | 0.8226 / 75.54 | focal_synth_w05 | no |
| equal_mean3 | 0.8781 / 83.07 | 0.8164 / 75.61 | w=0.333,0.333,0.333 | no |
| nnls_simplex | 0.8699 / 82.39 | 0.8113 / 75.15 | w=0.222,0.000,0.778 | yes |
| logistic_stack | 0.8520 / 81.47 | 0.7897 / 74.75 | coef=4.248,-0.202,-1.096 b=+0.836 | yes |
| **dirichlet_evid** | **0.8700 / 82.34** | **0.8084 / 75.23** | s=2980.958,0.061,804.823 | **SELECTED** |

Selected AVGC vs tip **0.8245/75.64**: **ΔH -0.0161 / ΔAcc -0.41**.

## Selected per-held

| Held | fit summary | H / Acc |
|------|-------------|---------|
| AVGC | s=2980.958,0.061,804.823 | 0.8084 / 75.23 |
| DTU | s=2980.958,0.036,1344.187 | 0.9016 / 88.85 |
| KUL | s=2980.958,0.080,4.727 | 0.8644 / 81.43 |
| UHD | s=2980.958,0.104,4.584 | 0.8526 / 80.08 |
| USTC | s=2980.958,0.040,967.541 | 0.9230 / 86.10 |

## External (bookkeeping; Acc tip not claimed)

| Variant | AASD H / Acc | Carta H / Acc |
|---------|--------------|---------------|
| focal_synth_w05 | 0.6208 / 51.06 | 0.5494 / 49.14 |
| **dirichlet_evid** | **0.6189 / 51.10** | 0.5411 / 47.92 |

## Targets

| Target | Result |
|--------|--------|
| AVGC ≥0.85 / Acc≥78 | **miss** (0.8084 / 75.23) |
| Tip improved (ΔH≥5e-4 vs 0.8245; Acc alone ≠ tip) | no / flat-hurt |

## Verdict

Selected **`dirichlet_evid`**: LODO mean **0.8700/82.34**, AVGC **0.8084/75.23** (Δtip -0.0161 H / -0.41 Acc). Does not clear AVGC 0.85/78. Tip not improved.

## Do not repeat

Do not re-grid train-only NNLS/simplex / logistic / Dirichlet-evidential stacks on frozen focal+synth+IRM tip stem logits at HMM 0.994/0.8. Distinct from (still closed) multi-stem soft-attn / T / isotonic on tip fuse.

## NEXT IDEA

Decision-layer stacking on frozen tip stems saturated/hurt vs tip. **CLOSED** stacking family. Prefer **group-dro-stem-avgc** (Group-DRO / CVaR worst-dataset stem training from tip inits; train-time logits) — not more PP on tip fuse: **no T / iso / stack / attn / agree**.

