The result
I started this project hoping to design a better p-tau217 sensor before anyone had to build one. I rebuilt a published carbon-nanotube transistor from its physics, then let optimizers loose on it. They found stronger signals almost immediately.
Then I checked where those signals lived. Most of them sat outside the range the real experiment had ever measured. Once I forced the search to stay inside the evidence, the improvement disappeared.
So the honest answer is a negative one. It is not proof that a better sensor cannot exist. It is a clear picture of how an optimizer gets better at the simulation than at the device, and I kept that result instead of hiding it.
NeuroNanoTwin began as an attempt to computationally improve a p-tau217 CNT-FET nanosensor. The study instead found that unconstrained optimization could exploit experimentally unsupported regions of the computational model. When support constraints and falsification audits were applied, no materially superior robust configuration was identified.
The Wang experiment only ever measured responses between 0 and 0.3281. That measured range is the envelope.
For every candidate the model predicts a response in 72 assay conditions: nine concentrations times eight ionic strengths. A prediction outside the envelope describes something the experiment never observed, so the model is extrapolating there. P_OOD is the share of the 72 that land outside. A configuration counts as supported when at most 14 of 72 do, which is P_OOD ≤ 20%. No configuration in the whole domain gets below 13 of 72.
…
Stage 4 grid · every dot is one modeled configuration · yellow is supported
of the unconstrained optimum's fresh conditions fell outside the envelopetable_7_constraint_ablation.csvStage 5 calibration median over 5,000 draws · 181 of 288, 62.85%, at nominal calibration · scale 1.5, 5.0 nm · map derived from the frozen fit
- 9,180,000B0, B1, B2 under calibration uncertainty on fresh conditions
- 26,049,762dense support and objective audit
- 32,130,000constraint-ablation calibration audit
- 10final fitted-model predictions at the Wang points
validation evaluations in the recorded 6.94 s, zero material improvementsstage5_summary.jsonmeaningful model-condition evaluations · 6.94 s wall time in the recorded environment · about 9.7 million per second, DERIVED