Overnight Proposals - 2026-07-03#

Recon Sources#

Candidate Scores#

candidate

novelty

feasible tonight

impact if true

leakage risk

decision

Leakage-safe forecastability-vs-horizon curve

4

5

4

2

chosen

Nuisance-invariant residual representation with probe gate

4

3

4

3

defer

Cross-modal missing-modality recoverability/identifiability gate

5

2

5

3

defer

Chosen Candidate#

Build M4: a forecastability-vs-horizon information curve. Labels are shifted within each patient only, then RFS is recomputed per horizon against the best nuisance/trivial baseline. This directly extends the existing RFS ladder and gives the repo a compact benchmark-method contribution without adding a new model family.

Why The Other Two Wait#

Nuisance-invariant representation learning is useful, but the current M3 artifact already shows the cheaper issue: the model must first beat moving-average/history baselines. Cross-modal recoverability is closer to the broad NeuroTwin v1 claim, but needs paired multimodal real data and stronger baseline coverage than this workspace can honestly create overnight.