Engine Cross-Check

Gate: Tucuxi reproduces the reference patient before any fit

The Colin 2019 covariate equations were written into a Tucuxi model file, which Tucuxi evaluates itself. Before any posterior comparison, Tucuxi’s prior-only prediction for the reference patient (35 y, 70 kg, SCr 0.83 mg/dL; 1000 mg q12h over 1.75 h) had to match the independent reference values; a mismatch would have stopped the run.

Reference values are the closed-form and matrix-exponential oracle values documented on the Literature Reproducibility page. Unlike the 30 May 2026 snapshot, the prior calculation itself was therefore compared, not only the fit.

Agreement between the two programs · n = 200

Difference between the two programs’ individual (posterior) estimates, Vancomyzer relative to Tucuxi, per parameter. Q and V₂ are reported without a criterion: two levels barely inform them, so both programs stay near the shared prior.

Acceptance criteria (fixed 17 Sep 2026, before the run)

• Clearance median |Δ| ≤ 2%0.84% met
• Clearance 95th percentile |Δ| ≤ 10%4.81% met
• Central volume median |Δ| ≤ 3%0.87% met
• Failed or excluded fits ≤ 2% of cohort0 of 200 met
• Every |Δ CL| > 10% explained individually0 such cases met

Steady-state exposure computed from each program’s own estimates: AUC₂₄ median |Δ| 0.85% (95th percentile 4.59%, maximum 8.43%), peak 0.57%, trough 1.54% (maximum 12.52%; the trough is the quantity most sensitive to clearance). Meeting these criteria means the two fitting implementations agree on this synthetic cohort; it does not mean the programs are interchangeable in practice.

Error against the known synthetic truth

Because the patients are simulated, their “true” parameters are known. Median absolute percentage error of each program’s estimates against that truth, with the unfitted prior for reference (lower is better).

As in the earlier snapshot, only clearance improves materially with a two-level fit (31.4% with the prior alone, 10.0% for Vancomyzer, 10.1% for Tucuxi). The synthetic truth comes from a different published model (Goti 2018) than the prior, so this is a check that the fit moves toward the truth, not a claim about accuracy in patients.

Largest disagreements — and why

The two programs weight a measured level differently. Vancomyzer’s assay-error model is σ = max(1, 0.15 × max(observed, predicted)); Tucuxi’s standard “mixed” model is σ = √((0.15 × predicted)² + 1²). They are close when a level is near its prediction and diverge when a level is well above it. To test whether that explains the tail, every patient with a clearance difference of 3% or more (14 of 200) was re-fitted by a third, independent estimator (numpy/scipy, sharing no code with either program) under each error model in turn.

The independent refit reproduces each program to within 0.01% under its own error model, and the error-model form alone accounts for the whole difference in every case (largest: p101, 9.22%). No optimiser, convergence, boundary or model-integration difference was found. Unlike the 30 May 2026 snapshot, no disagreement is left unexplained. Whether Vancomyzer’s error model — which gives a little less weight to a level that comes back unexpectedly high — is the preferable choice is a clinical design question that this analysis records rather than settles.

Method

1. Generate 200 synthetic ICU patients (seed 42) with “true” parameters drawn from a Goti 2018–based model; simulate two levels with assay error (15 mg/kg (rounded to the fixture dose) q12h, 1.5 h infusions, steady state; levels at 3.0 h and 11.5 h after the start of the dose). The fixture is committed.
2. Run Vancomyzer’s engine (manifest 2026-09-17.1) on each patient: Colin 2019 prior from the covariates, then the Bayesian fit on the two levels.
3. Write the Colin 2019 equations into a Tucuxi model file (vancomyzer.crosscheck.colin2019.tdd, SHA-256 49f6ebe6bcb31c79…) with prior variability (log-scale SD: CL 0.35, V₁ 0.25, Q 0.5, V₂ 0.5) and Tucuxi’s mixed residual error (1.0 mg/L additive, 15% proportional). Check its prior-only prediction against the reference patient (table above). Then give Tucuxi the same dosing history and the same two levels per patient and run its Bayesian fit.
4. Score the two sets of estimates with the committed comparison script against the criteria written down on 17 Sep 2026, before the run; re-fit every tail case independently under both error models.

Comparator: Tucuxi-core (tucucli) (github.com/sotalya/tucuxi-core, commit d09737e3), built from source. The fixture, model file, query generator, raw result files, comparison and attribution scripts are in the repository at src/lib/validation/crosscheck/.

Scope & limitations

• Checks the fitting implementation, not clinical accuracy

  Both programs used the same Colin 2019 prior (Tucuxi evaluating the equations from its own model file) and the same simulated data. Agreement shows that two independently written Bayesian estimators reach the same answer on the same problem. It does not show that Colin 2019 is the right model for any population, and it says nothing about accuracy in patients.

• Error models differ by design

  The two programs' residual error models are not identical in form, and this difference accounts for the entire tail of the distribution. The comparison therefore also functions as a sensitivity analysis of that choice: up to about 9% in clearance for individual synthetic patients whose level came back well above prediction.

• Simulated patients only

  The patients and their levels are computer-generated. Real-patient performance has not been evaluated; the Predictive Performance analysis is also synthetic.

• Two-level sampling

  A post-infusion level and a trough at steady state mainly inform clearance. Agreement on Q and V₂ largely reflects both programs staying near the shared prior.

• Open-source comparator only

  Commercial Bayesian dosing products were not run; no results for them are shown or implied.

What this shows: given the same prior and the same simulated data, Vancomyzer and Tucuxi produce individual estimates that agree within the pre-set criteria for all 200 synthetic patients, with every larger difference explained. What it does not show: that the Colin 2019 equations are clinically correct, that either error model is the right one, or that Vancomyzer is accurate in real patients.

Agreement between the two programs · n = 200

Relative difference between the two programs’ individual (posterior) estimates, as a percentage of the mean of the two, per parameter. Both were given the same prior and the same two simulated levels.

Observed agreement: median absolute differences of 0.23–0.85% across the four parameters, 95th percentiles of 2.93–5.65%, and maximum differences of 13.77–32.76% (clearance: median 0.81%, 95th percentile 5.65%, maximum 18.10%; central volume: 0.85%, 3.00% and 32.76%). No equivalence margins were set before the run, so these numbers describe the agreement that was observed; they are not a test of whether the programs are interchangeable.

Error against the known synthetic truth

Because the patients are simulated, their “true” parameters are known. Median absolute percentage error of each program’s estimates against that truth, with the unfitted prior for reference (lower is better).

The two programs’ median errors differed by 0.1–0.4 percentage points per parameter. No equivalence margins were prespecified. Only clearance improved materially with the two-level fit (31.7% with the prior alone, 10.2% for Vancomyzer and 10.3% for Tucuxi). The inter-compartmental clearance and peripheral volume (Q, V₂) barely moved: a steady-state peak and trough carry little information about them, so both programs stayed near the shared prior. Similar errors are expected when two programs fit the same model to the same data; this is synthetic truth, not accuracy in patients.

Largest disagreements

The five largest differences recorded for clearance and central volume (patient id, creatinine clearance in mL/min, difference).

The largest disagreements are not confined to patients with augmented renal clearance. Creatinine clearance for the patients listed ranges from 42 to 177 mL/min, and 2 of the 8 patients listed are above 130 mL/min. Listed patients with a BMI of 40 or more: p87 (BMI 41.0) and p196 (BMI 42.2). These cases belong to the archived software version. Differences in prior settings between the compared implementations were not fully recorded, so attribution remains unresolved. These individual disagreements have not been explained and should be investigated.

Method

1. Generate 200 synthetic ICU patients (seed 42, the same generator as the Predictive Performance page), with “true” parameters drawn from a Goti 2018–based model.
2. Simulate a steady-state peak and trough for each patient, with simulated residual error.
3. Compute per-patient priors using the archived software version, then run that version’s Bayesian fit on the two levels to estimate CL, V₁, Q and V₂.
4. Give Tucuxi a model file written by Vancomyzer (two-compartment, IV infusion) containing Vancomyzer’s prior values for that patient, with prior variability (log-scale SD: CL 0.35, V₁ 0.25, Q 0.5, V₂ 0.5) and residual error (mixed: 1.0 mg/L additive and 15% proportional) chosen to approximate Vancomyzer’s settings, plus the same dosing history and the same two levels. Run Tucuxi’s Bayesian fit.
5. Compare the two programs’ estimates parameter by parameter. The analysis checked that all 200 patients were present before summarizing.

Comparator: Tucuxi (github.com/sotalya/tucuxi-core, commit d36cc10 (2026-05-28)), an open-source dosing program developed by the REDS institute at HEIG-VD, Switzerland. The one-off scripts used for the run, and the per-patient inputs, were not committed to the repository.

Scope & limitations

• Checks the fitting implementation, not the Colin equations or clinical accuracy

  Both programs were given the same priors and a Colin model file written by Vancomyzer (a Colin 2019 file that Tucuxi can load was not publicly available). Agreement therefore shows that the two programs fit the same model to the same data in a similar way. It does not independently confirm the Colin 2019 equations (those are checked against published values on the Literature Reproducibility page), and it says nothing about accuracy in patients.

• Same starting estimate

  Tucuxi was given Vancomyzer's starting estimate for each patient, so the prior calculation itself was not compared.

• Simulated patients only

  The patients and their levels are computer-generated. Real-patient performance has not been evaluated; the Predictive Performance analysis is also synthetic.

• Two-level sampling

  A steady-state peak and trough mainly inform clearance. Agreement on Q and V₂ largely reflects both programs staying near the shared starting estimate.

• Run date and engine version

  Run once on 30 May 2026 with the Vancomyzer engine snapshot dated 30 May 2026; not the current release. It has not been repeated with the current engine.

What this shows: given the same priors, model file and simulated data, Tucuxi and Vancomyzer produced similar individual estimates for most of the 200 synthetic patients, with some larger individual differences that have not been explained. What it does not show: that the Colin 2019 equations are correct, that the two programs are interchangeable, or that Vancomyzer is accurate in real patients.