Adversarial review. C1 major_issues; C2 minor_issues; C3 minor_issues. Significance minor for each narrowly stated simulation finding; this is a small uncalibrated model, not a demonstrated clinical decision rule. I inspected the complete bundle and reran the original computations separately: all supplied R1-R3 outputs reproduce exactly. The attached instrumented run preserves the event generator, seeds, rates and stopping rules while adding outcome flags and final resistant counts. This is an audit of the supplied simulator, not independent clinical validation.

C1 contains a decisive unsupported assertion. The same500 adaptive seeds give186 observed progressions(37.2%) and314 living nonprogressed runs censored at300days(62.8%). There are zero cures, but236 adaptive runs have zero resistant cells at their endpoint. The continuous seeds give483 cures and17 progressions. Thus the two cure percentages reproduce, but all adaptive trials progressing and all establishing a resistant lineage are false for the observed simulated horizon. Code returns ttp300 for censored runs and main appends every noncure as a progression;260.6 is the average of event time capped at300 for adaptive runs, not the unrestricted mean progression time. Separate cure, progression and administrative censoring, report cumulative incidences with an explicit horizon and uncertainty, and relabel or reanalyse the time summary. Long-term inevitability is a different proposition requiring a proof or justified extended simulation; it cannot be inferred from zero cure. No conclusion about individual patients follows.

The max_days300 cap is absent from the paper's stated regimen stopping description, which says until eradication or progression. It is consequential, not a cosmetic implementation detail. Event time can also overshoot300 and be applied after the observation window; truncate the last wait and drug integral consistently. The dose120.4days averages exposure through progression OR censoring; comparison with57.1 continuous days uses different clinical endpoints and does not quantify toxicity.

C2's narrow model summary reproduces:68.6 versus74.8days,56.2 adaptive exposure versus68.6 continuous,18.1% reduction. Specify the fixed parameters,500 seeds, simulation horizon and sampling intervals; these values are not a general property of pre-existing resistance. Time to model progression is not survival, and days of exposure are not a clinical toxicity measure. Resistant cells have positive death rate so extinction is possible even when initially present: calling complete eradication impossible is mathematically too strong in this finite stochastic model, even if none occurred among these sampled seeds. A controlled factorial sensitivity over initial resistance, mutation rate, carrying capacity, threshold, fitness cost and treatment policy is needed before generalizing the mechanism. In particular, absence of detected mutations does not mean true zero resistance, and this study cannot require a clinical biopsy strategy or establish optimal treatment.

C3's endpoint ratios1.11 and1.09 reproduce. Smaller Euler steps(.025,.005 versus.05) leave these rounded ratios unchanged; attached convergence results support that narrow numerical conclusion. But the cost sweep is nonmonotonic: at.10 the ratio is1.12, above.20's1.11; at.05 it is1.09. Do not describe a systematic monotonic contraction. Crowding alone yields modest delay in this particular parameterization and AT50 policy, not across all competitive systems. With zero fitness cost, differential drug mortality and mutation can still change clonal composition, so the paper's statement that crowding cannot alter underlying clonal composition should be stated in its restricted off-drug, mutation-free context, if intended.

The abstract and Discussion leap from a single synthetic model to adaptive therapy requiring pre-existing resistance for clinical benefit and aggressive upfront eradication being optimal otherwise. Neither conclusion is established by comparing two policies at one main parameter set; no empirical tumor calibration, clinical outcomes, detection-error model, toxicity function or optimization was performed. Remove clinical directives and frame as a model-specific hypothesis. This review does not recommend changes in cancer care.

Numerical reproduction is compatible with these major interpretive issues: the reproduced outputs lack the endpoint distinctions needed for C1. The attached event counts provide a direct falsifying check of its all-trials wording. No author identity was sought or learned.