LONGITUDINAL PERSONALIZATION + LOCKED VALIDATION

From target percept to a patient-specific constrained cortical policy.

Version 0.2 separates calibration episodes from held-out validation, updates a persistent patient model, models multi-electrode interaction, and issues a replayable evidence certificate.

Target percept → patient cortical model → device constraints → stimulation policy → frozen oracle → ACT / ABSTAIN / FALLBACK

Patient calibration

Patient profileTraining episodesValidation episodesElectrode dropout %

Model fidelity RMSE0.041

Held-out target RMSE0.184

Task utility0.221

Unsafe ACT rate22.2%

Learned patient parameters

Parameter	Before	After	Status

Genesis paired tests

Test	Snapshot	AVCR v0.2
Same present / different history	Collapsed	Separated
Gaze remapping	World drift	World-stable
Electrode loss	Stale plan	Fallback → remap
Unknown state	Overconfident ACT	ABSTAIN
Multi-electrode interaction	Independent dots	Interaction-aware

Benchmark design

The compiler does not assign its own final score. A frozen nonlinear oracle with perturbed geometry, anisotropy, saturation, adaptation and pairwise interactions evaluates all policies on held-out seeds.

Decision gate: WAITING

From target percept to a patient-specific constrained cortical policy.

Patient calibration

Learned patient parameters

Genesis paired tests

Benchmark design

Replayable evidence certificate