# results_isotonic_ds_fuse

**Claim:** `isotonic-ds-fuse` — train-only **isotonic regression** (monotone map tip fuse scores → calibrated probs → logit) on frozen **focal↔synth fuse w=0.5**. Per-dataset or global fit on non-held; leave-AVGC fits without AVGC labels.

**Distinct from** `causal-ds-temp` (temperature / T grids — CLOSED), `platt-acc-tip` (isotonic on Acc SW2 streams), `focal-synth-adapt-readout` (per-subj EWMA), multi-stem soft-attn pool.

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

**Compute:** CPU; reuse `models/focal_synth_adapt_readout/cache/`. Code: `isotonic_ds_fuse.py`. JSON: `results/research/isotonic_ds_fuse.json`.

## Selection

| Rule | Pick |
|------|------|
| mean LODO Acc then H (isotonic variants) | **`held_iso_per_ds_pick`** |
| Tip to beat (held_T_nll_med) | **0.8245 / 75.64** |
| Raw fuse sanity | **0.8226 / 75.54** |

Ranked Acc: held_iso_per_ds_pick → held_iso_global → global_iso_all → per_ds_iso_emis_gheld

## LODO (preset)

| Variant | mean H / Acc | AVGC H / Acc | iso_src@AVGC | selectable |
|---------|--------------|--------------|--------------|------------|
| raw | 0.8821 / 83.36 | 0.8226 / 75.54 | identity | no |
| global_iso_all | 0.8751 / 83.92 | 0.8109 / 76.62 | pooled_train | yes |
| held_iso_global | 0.8762 / 83.94 | 0.8121 / 76.59 | pooled_train→held | yes |
| **held_iso_per_ds_pick** | **0.8761 / 83.95** | **0.8123 / 76.60** | KUL | **SELECTED** |
| per_ds_iso_emis_gheld | 0.8751 / 83.92 | 0.8120 / 76.61 | per_ds_emis+pooled_held | yes |

Selected AVGC vs tip **0.8245/75.64**: **ΔH -0.0122 / ΔAcc +0.96**.

## Selected per-held

| Held | iso_src | H / Acc |
|------|---------|---------|
| AVGC | KUL | 0.8123 / 76.60 |
| DTU | AVGC | 0.9173 / 90.82 |
| KUL | AVGC | 0.8774 / 83.65 |
| UHD | KUL | 0.8589 / 81.80 |
| USTC | AVGC | 0.9145 / 86.85 |

## External (bookkeeping; Acc tip not claimed)

| Variant | AASD H / Acc | Carta H / Acc |
|---------|--------------|---------------|
| raw | 0.6208 / 51.06 | 0.5494 / 49.14 |
| **held_iso_per_ds_pick** | **0.6256 / 51.03** | 0.5570 / 49.54 |

## Targets

| Target | Result |
|--------|--------|
| AVGC ≥0.85 / Acc≥78 | **miss** (0.8123 / 76.60) |
| Tip improved (ΔH≥5e-4 vs 0.8245; Acc alone ≠ tip) | **no** (H −0.0122; Acc +0.96 tradeoff) |

## Verdict

Selected **`held_iso_per_ds_pick`**: LODO mean **0.8761/83.95**, AVGC **0.8123/76.60** (Δtip −0.0122 H / +0.96 Acc). Does not clear AVGC 0.85/78. **Tip not improved** (H regresses; do not adopt Acc-only gain).

## Do not repeat

Do not re-grid train-only isotonic (global / held-only / per-ds nested-pick / per-ds-emis) on frozen focal↔synth fuse w=0.5 at HMM 0.994/0.8. Still do not reopen causal-ds-temp T grids.

## NEXT IDEA

Isotonic on tip fuse saturated vs tip. Prefer **evidential-tip-stack** (Dirichlet / non-neg or logistic stack over frozen tip stems focal/synth/IRM; train-only causal linear/isotonic/logistic on logits — ≠ closed multi-stem soft-attn pool) — not Mixup/SSL/DANN/T/isotonic.

