# results_causal_ds_temp

**Claim:** `causal-ds-temp` — causal HMM posterior smoothing on frozen **focal↔synth fuse w=0.5** evidence. The posterior stays soft; scoring labels remain binary at P(B)>0.5.

**Method:** causal EMA on the two-state HMM posterior. Tau is selected inside each outer LODO fold using only its four training datasets; tau=0 is the exact existing HMM baseline.

**Honesty:** LODO synthetic-switch labels select tau by inner Acc then H at HMM **0.994/0.8**. External tau is selected by 5-fold corpus validation. AASD/Carta labels are used only for final scoring.

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

## LODO

**Selected:** `hmm_posterior_ema` — mean **0.8821/83.36**; raw fuse AVGC **0.8226/75.54**.

| Outer held-out | tau (s) | H | Acc % | L | SDT (s) | undetected / switches |
|---|---:|---:|---:|---:|---:|---:|
| AVGC | 0.00 | 0.8226 | 75.54 | 0.903 | 7.20 | 178 / 890 |
| DTU | 0.00 | 0.9207 | 90.44 | 0.938 | 2.20 | 32 / 220 |
| KUL | 0.00 | 0.8783 | 82.86 | 0.934 | 3.33 | 153 / 934 |
| UHD | 0.00 | 0.8587 | 80.69 | 0.918 | 1.60 | 116 / 661 |
| USTC | 0.00 | 0.9300 | 87.25 | 0.996 | 0.19 | 26 / 408 |
| **mean** | — | **0.8821** | **83.36** | — | — | — |

### Legacy temperature controls (not selected)

| Variant | mean H / Acc | AVGC H / Acc | selectable in this run |
|---|---:|---:|---|
| `raw` | 0.8821 / 83.36 | 0.8226 / 75.54 | no |
| `global_T_nll` | 0.8821 / 83.36 | 0.8226 / 75.54 | no |
| `held_T_nll_mean` | 0.8844 / 83.37 | 0.8246 / 75.64 | no |
| `held_T_nll_med` | 0.8844 / 83.38 | 0.8245 / 75.64 | no |
| `held_T_nll_pick` | 0.8821 / 83.36 | 0.8226 / 75.54 | no |
| `held_T_causal_grid` | 0.8816 / 83.32 | 0.8210 / 75.44 | no |
| `per_ds_emis_T_mean` | 0.8815 / 83.35 | 0.8202 / 75.52 | no |

Selected AVGC vs tip **0.8226/75.54**: **ΔH +0.0000 / ΔAcc +0.00**.


## External real-switch evaluation

Tau **0.00 s**, selected on the five corpus datasets only; external labels were not used for selection.

| Variant | AASD H / Acc | L / SDT (s) | Miss rate | Carta H / Acc | L / SDT (s) | Miss rate |
|---|---:|---:|---:|---:|---:|---:|
| `raw` | 0.6208 / 51.06 | 0.792 / 12.66 | 0.348 | 0.5494 / 49.14 | 0.623 / 26.69 | 0.259 |
| **`hmm_posterior_ema`** | 0.6208 / 51.06 | 0.792 / 12.66 | 0.348 | 0.5494 / 49.14 | 0.623 / 26.69 | 0.259 |

## Targets

| Target | Result |
|--------|--------|
| AVGC ≥0.85 / Acc≥78 | **miss** (0.8226 / 75.54) |
| Tip moved (\|ΔH\|≥5e-4 or \|ΔAcc\|≥0.05) | no / flat-hurt |

## Verdict

Selected **`hmm_posterior_ema`**: LODO mean **0.8821/83.36**, AVGC **0.8226/75.54** (Δtip +0.0000 H / +0.00 Acc). Does not clear AVGC 0.85/78. Tip not improved.

## Interpretation

The posterior EMA changes decision persistence only; it does not create new EEG evidence. A smoother probability trace is useful only if it also improves held-out switch delay/miss rate without unacceptable Acc loss. The five LODO corpus evaluations use synthetic switch schedules, not genuine neural switch events.

