Simulation and validation status

Study design

The stationary robustness grid contained 108 scenarios: true MT 20–90% MSO; true response spreads of 3%, 7% or 15% of MT; 0%, 2% or 5% random-response mixtures; and Gaussian or variance-matched logistic response curves. A further 19 scenarios tested drift, incorrect initial responses, ceilings, starting intensities and output boundaries.

The unchanged production estimator used a default first intensity of 37% MSO and fixed 7% model spread. Synthetic response probabilities were generated independently of the estimator. The paired plans shared their first 20 random draws; early stops were retained. Computational anchors were never counted as delivered pulses.

20 versus 30 pulses

Gaussian response curve, true spread 7% of MT, no random-response mixture.

At 30 pulses, the Monte Carlo 95% interval for the within-5-point rate was 99.84%–99.99%. Monte Carlo intervals describe repeated-simulation precision for this design; they are not patient-level MT confidence intervals. Ten extra pulses reduced robustness-grid MAE by 0.301 MSO points (paired interval 0.296–0.306).

Accuracy across every stationary scenario

Explore a recorded simulation

Use the pulse slider to inspect delivered intensity, response and the evolving estimate. These are fixed synthetic study records; viewing them does not add observations to your assessment.

Blue line: estimated MT · dashed line: true MT · green circles: positive responses · brown squares: negative responses.

This is the largest observed review-ready error, not a representative case. True MT 90; Gaussian response spread 15%; random-response mixture 5%. An early positive at 37% MSO redirected sampling downwards. Two positives in the last ten pulses allowed the run to pass the 0/1/9/10 response-imbalance rule.

Stress-test results

Stress scenarios were analyzed separately and should not be pooled into stationary accuracy. The table shows 30-pulse plans; the report includes each of the 19 scenarios and both budgets.

A stationary model lagged abrupt threshold changes. Wrong initial responses caused persistent errors but all 2,000 such runs were rejected by the readiness gate. All 2,000 ceiling-limited runs stopped after one pulse; no above-ceiling pulse was simulated. Safe stopping can leave a reachable threshold unresolved.

What the MTAT checks establish

37 of 37 isolated scalar reference cases matched. These retained numerical checks used original MTAT 2.1 scalar routines with environment stubs. They did not execute the full MTAT application or verify its interface, initialization pathways, stopping rules or readiness behavior end to end.

The 127,000-history simulation evaluated CTDS against known synthetic thresholds. It was not a head-to-head performance comparison with MTAT. Numerical agreement in a limited set of cases does not establish similar clinical performance or full-application equivalence.

Independent clinical validation is pending

Further work requires stronger estimate-quality checks, evaluation on held-out scenarios, complete original-application comparisons and a prospective clinician-supervised EMG study against an independent reference assessment. No clinical acceptance margin was established here. Visual-movement, active-MT and custom-spread performance were not validated.

Full report and reproducible study

Study conducted on website version 42. The estimator is unchanged in this publication update. The package includes the frozen code, analysis plan, seeds, raw endpoints, scenario summaries and pulse replays. Nine simulation checks passed; the retained 37 scalar checks also matched.