Reviews and checks

Verdict: minor_issues. The registered test is run as planned and the reported non-rejection is sound; the wording claims linearity the test does not establish, the survey design is not used, and relevant prior work is missing.

Significance: minor. A dimensional structure of depression is the prevailing finding in the taxometric literature; showing it with physiological and functional load measures in a population sample is a small, new step.

C2: no stronger somatic coupling in the highly impaired

What holds. Partial Spearman correlations of load-bearing limitation with somatic symptoms (adjusted for cognitive-affective symptoms, age, sex) are 0.161 in 454 highly impaired adults and 0.175 in 9,291 others; the bootstrap difference is -0.014 (95% CI -0.104 to 0.077), one-sided p 0.62. This matches the plan.

Issues.

1. Non-rejection, not equivalence. The CI admits a difference up to +0.077, about half the size of either correlation, so "no stronger" is supported only in the sense that no increase was detected. No smallest effect of interest was registered.
2. Strata defined on a severity-linked item. PHQ item 10 is strongly related to total score, so the high-impairment stratum has a restricted, higher range of somatic and cognitive scores. Range restriction changes correlations within strata independently of any true change in coupling; the comparison is not of like with like. The paper notes the dependence in Limitations, but the claim does not carry the caveat.
3. Covariates. Unlike C1, H2 omits survey cycle and age squared, and treats binary sex by rank; minor, but it differs from the model used elsewhere.
4. Unweighted, as above.

Verdict: minor_issues. The estimate is as registered and reported correctly; the claim reads as stronger than a non-significant difference with a wide interval in a range-restricted subgroup supports.

Significance: minor. It is a secondary, exploratory test; knowing the coupling is not detectably stronger in impaired adults is of limited use given the interval width.

With it in its evidence: verdicts.json

What supports it:

• The design is a sound, pre-registered Hansen-type sup-t test: a hinge term max(0, PHQ - c) searched over a grid, with the null distribution from a wild bootstrap of the linear model that recomputes the whole search, so the grid search is accounted for. A negative control (height) is included and shows no break (raw p 0.478).
• Raw one-sided p-values (struct 0.409, BMI 0.959, allostatic 0.999) give Holm values of 1.0 for all three, which is computed correctly by holm().

Issues:

1. The claim says more than the test checks. The test is one-sided: it asks only whether the slope steepens above some c. The claim and title say the measures "rise linearly ... across the full range". For allostatic load the maximum t over the whole grid is negative (sup_t = -1.723), i.e. at every c from 8 to 22 the estimated slope change is negative, and at the best break (c = 22) slope_below 0.026 SD plus change -0.085 SD gives an estimated slope above 22 of about -0.06 SD per point; BMI is similar (0.021 - 0.049). That is a departure from linearity in the other direction (concavity/flattening), which the one-sided statistic cannot detect and no two-sided or curvature test assesses. The supported statement is "no evidence of a steeper slope above any threshold from 8 to 22", not "linear". Suggest rewording C1 and the title, or adding a two-sided test (or a test of any departure from linearity) as an exploratory analysis labelled as such.
2. Power for a negative result is not shown. For a negative claim the review must ask whether the test could have found the effect. Only 120 adults score 20-27 and 345 score 15-19, so slope changes above about 18-22 are estimated from very few people. No minimum detectable slope change, simulation of the test under a planted break, or confidence interval for the slope change at each c is reported. Without it, "no threshold" cannot be distinguished from "a threshold the data are too sparse to see" in the severe range, which is exactly the range the hypothesis concerns. Suggest a simulation: plant breaks of given size at c = 15, 18, 20 and report detection rates at this n.
3. Uncertainty and effect sizes missing in Results. Linear slopes and slope changes are given as point estimates in SD per point with no standard error or CI (the style guide asks every estimate to carry its uncertainty, declared as a result). The Summary reports only Holm p = 1.0 three times, which carries little information; the raw p-values and the slope-change CIs would.
4. Survey design ignored. NHANES is a complex, weighted sample; the analysis is unweighted and ignores strata/PSUs. The paper's "in this population" reads as the US population; unweighted estimates describe the analysed sample. Either state this in Limitations or use the MEC weights and design.
5. Outcome scale. Load-bearing limitation is a 0-3 mean with many exact zeros (screened-out respondents are set to 0 in structural()), so a floor at low PHQ and a ceiling near 3 can bend the curve independently of any state change; the Discussion mentions floor/ceiling, which is good, but the linear Gaussian model's t-statistics on such an outcome rely on the bootstrap; this is fine as done (wild bootstrap is heteroskedasticity-robust in distribution) but should be stated.

Repeatability: good. Data are pinned by URL, size and SHA-256; code, seed and package versions (numpy 2.5.3, scipy 1.18.1, pandas 3.0.6) are given; sh code/run is the entry point. Missing: a Python patch version / container definition (only requirements.txt), and an explicit statement in Methods of how missing outcome values are handled per measure (allostatic requires >= 7 of 9 markers, quartile cut-offs computed within the analysed sample; these are in code/nhanes.py but not in Methods). The code/nhanes.py docstring still says "Usage: python code/analyze.py", a harmless leftover from the earlier analysis.

Significance: minor. A careful cross-sectional negative, but as the paper itself says, population averages cannot test within-person bistability, so it bears only weakly on the attractor-state hypothesis.

C2 (negative result: coupling not stronger in highly impaired) — minor_issues

What supports it: pre-registered, the partial Spearman correlations (0.1607, n = 454 vs 0.1746, n = 9,291) and bootstrap CI of the difference (-0.1036 to 0.0773) are computed as the plan says; the CI bounds the possible strengthening to below about 0.08, which is an informative negative.

Issues:

1. Range restriction / collider-like stratification. Strata are defined by PHQ item 10, which is strongly related to symptom severity. Within the high-impairment stratum the somatic and cognitive scores have a different, compressed range, which by itself changes correlations; a difference (or its absence) between strata is therefore not a clean test of "coupling strengthens". The Limitations mention that the item relates to severity, but the consequence for C2 should be stated, or the analysis repeated with an interaction of item 10 with somatic score in the full sample, adjusting for total severity.
2. The one-sided bootstrap p (0.615) named in falsified_if is not reported in the claim or the Results.
3. H2 adjusts for age and sex but not survey cycle (H1 adjusts for cycle); minor inconsistency, matches the plan.
4. Item 10 is asked only when any symptom is present, so adults with PHQ = 0 are excluded from both strata (9,745 of 14,722 analysed); worth stating in Methods.

Significance: minor.

Style-guide departures

• Estimates in Results lack uncertainties (SE/CI) and are not declared with them.
• Title states "rise linearly", which the test does not establish (see C1.1).
• Otherwise the sections, placeholders, citations and provenance follow the guide.

With it in its evidence: verdicts.json

The supplied program was rerun in an offline, restricted Docker container using Python 3.12.15 and the specified NumPy 2.5.3, SciPy 1.18.1 and pandas 3.0.6. Separate independent regression calculations use Frisch-Waugh residualization and QR projections; they reuse the author's measure-construction loader, so they are an independent check of the model fitting, not an independent validation of every questionnaire coding decision. Their full threshold grid and aggregate results accompany this report. The diagnostic intervals are ordinary pointwise model-based intervals, not multiplicity-adjusted or survey-design-corrected inference.

The rerun matched the complete R1 output exactly, including all 16 declared numerical evidence bindings. Independent partial-rank estimates also agree (0.160680 versus 0.174553). These checks support implementation agreement, not the stronger absence claims.

C1: major issues

A one-sided non-rejection does not establish linearity or absence of thresholds. The program searches for a positive hinge coefficient, maximizing its signed t statistic over PHQ thresholds 8 through 22. It rejects only sufficiently strong steepening. It is not an omnibus nonlinearity test, does not test flattening, and provides no equivalence tolerance or simultaneous upper bound on slope changes. A relation with a large negative hinge is nonlinear but is not the alternative this test targets. Allostatic and BMI maxima are already negative in the reported output; large positive-tail p-values are therefore unsurprising. Holm-adjusted p-values of 1 are not probabilities that the linear hypothesis is true.

The narrow supported account is that the registered procedure did not detect an upward change in slope in these unweighted complete-case samples. Replace 'rise linearly across its full range', 'no threshold', and 'one continuum, not two states' with that statement. Standing height's non-rejection likewise does not show that it has no breakpoint or validate the test's ability to detect meaningful changes.

The reported break is not the best-fitting breakpoint. In sup_t, the selected threshold maximizes signed t for upward curvature; it does not minimize residual sum of squares or maximize likelihood across two-sided piecewise models. Calling this the 'best-fitting break' is especially misleading when all coefficients are negative: the procedure selects the least compelling upward-change candidate, not the strongest flattening. The independent grid diagnostic explicitly compares the signed-t maximizer with the minimum-SSE fit. For allostatic load the reported signed-t maximizer is 22 (t = -1.723), whereas the minimum-SSE candidate is 8 (t = -3.007); for BMI these are 21 (t = -1.277) versus 8 (t = -4.112). Only 46 allostatic observations lie above 22 and 65 BMI observations above 21. This is direct evidence that the top-of-scale candidates are not the best-fitting breaks. These exploratory t values are not search-adjusted evidence establishing a new threshold. Any interpretation of those selected estimates as evidence for a top-of-scale plateau needs a separately identified exploratory analysis and selection-aware uncertainty.

Sensitivity and population scope remain unresolved. Only 120 of 14,722 complete-PHQ adults have scores 20-27, and outcome-specific numbers are smaller. Report counts above each candidate threshold and detectable or excluded slope changes. The recorded positive slope coefficients of linear models do not test whether those models adequately describe the full range. The proposed within-person attractor interpretation also does not entail a common cross-sectional hinge: heterogeneous individual thresholds can average smoothly. The paper acknowledges this in its Discussion, but its headline and Summary still contrast one versus two states more strongly than the design permits.

The survey design is omitted. nhanes.load selects only SEQN, age, sex and income from DEMO and never retains survey weights, strata or PSUs. The fit is unweighted OLS; wild-bootstrap signs are independent by person, and the stratum analysis resamples people independently. Adjustment for age, sex and cycle is not a substitute for the sampling design. The resulting estimates can be described as unweighted sample analyses, but their reported uncertainty is not design-based inference about US adults. NCHS explains the roles of weighting and clustering in its variance guidance and weighting guidance. Either narrow the target explicitly or supply an appropriately weighted, survey-aware sensitivity analysis; this review does not assume which direction such a correction would move the estimates.

C2: major issues

The point estimates, 0.1607 versus 0.1746, describe a slightly smaller association in the high-impairment sample. The declared interval for high minus lower impairment is -0.1036 to +0.0773. It includes both weaker and stronger associations. A positive difference of 0.05, for example, is compatible with that interval. Thus 'is no stronger' is not established; use 'no strengthening detected', give the interval and avoid equivalence language without a justified margin.

There are only 454 highly impaired respondents versus 9,291 in the lower stratum. Functional difficulty is related to symptom severity and physical functioning, so stratification can also alter associations by selection or restricted ranges; it does not isolate an independently assigned severe-depression state. Partial rank correlation after adjustment for the cognitive-affective score does not establish causal bodily coupling. The independent-person bootstrap shares the survey-design limitation above. Preserve the reported descriptive contrast, but do not infer absent strengthening or a single underlying dynamical state.

Required revision

Retain the registered directional test and all its null results. Identify it as a test for detectable upward hinges, label its selected threshold correctly, report uncertainty and tail counts, distinguish unweighted sample results from population inference, and limit the impairment conclusion to the displayed confidence interval. A two-sided curvature test, survey-aware analysis or equivalence design would be additional work and should be disclosed separately. Registration and exact reproduction protect the computational record but cannot expand the hypothesis that was tested.

With it in its evidence: binding_checks.json, diagnostic.json, diagnostic.py, input_check.json, reproduced_R1.json, reproduction.log

Verdicts

Claim IDs: C1 is claim:ded2d5480652cf7e36b4fa12d66fb0d7ce3a1fea3ec4b34da1c290ef4dabcfa8; C2 is claim:59c4ad2dbb810e015aa585ca002109e98b2bffc70ef0b77f7646119407bed7aa.

Reported over budget: the work took more than the minutes its bundle declares.

Results

A number agrees when it lands within its tolerance of the declared value, compared as the decimals canonical JSON writes; anything else must be equal.

Hidden content

Before any model read the bundle, the harness's scan found nothing hidden in its 17 text files.

Files

• No run.log: the run printed nothing.
• environment.json: the machine, engine, image, command, limits, and outcome.
• The run wrote nothing under results/.

With it in its evidence: environment.json