Hard-stop review#

This page records the review of the inherited hard stops in the future investigations grid of the roadmap. The review applies the Eligibility rule to each grid row. It splits each row into sub-claims and gives each sub-claim one verdict. A sub-claim moves to a main-roadmap row when an established method or a natural extension carries it. A sub-claim stays in the research queue when it needs an argument of a new kind.

A verdict is not a capability. “Moves” means that the owning row may implement the part. The part ships only when that row delivers its contract, its nonzero witnesses, and its registered study. Until then the refusal stays, and it cites the owning row. The review read the code at main commit bb7a0ce.

The terms#

term	meaning
sub-claim	one request that a grid row refused, stated as one target, one estimator and one inference claim
refusal today	what the package does for the sub-claim. A quoted message gives its source location in `src/cleverly`
sources read	the sources the review read first-hand, with the version and the exact locator. References records each first-hand read with its locators. An abstract-only read carries no verdict weight, and References can omit it
base result and steps	the published result that the extension starts from, and each step that carries it. Each step names its row in the standard-step table of Eligibility
objection	a source that records a step as open or as a defect, or a step that needs an argument of a new kind
verdict	“moves” with its acceptance route, “keeps” with the kind of missing argument, “no action”, “not adopted”, or “not a claim of this row”
owning row	the main-roadmap row that implements a moved part, or the research-queue row that keeps the claim

The research queue uses six kinds of missing argument.

kind	the argument that is missing
identification	a law or an assumption that identifies the target from the observed data
regularity	a pathwise derivative of the target, or a model in which the estimator is regular
asymptotic	a limit theorem for the estimator as the package builds it, such as a remainder bound or a bootstrap consistency result
adaptive-optimum inference	a limit law after a data-adaptive selection, or for a target with no fixed limit
arbitrary dependence	an argument for a fixed number of dependent units
serial dependence	a limit theorem for split-based TMLE under mixing

The verdicts#

Simulated confounding#

The base law of every row in this group is the pinned DoWhy refuter. The review read dowhy/causal_refuters/add_unobserved_common_cause.py at commit 2116d5c. The page cites its functions by name, because two counts of its lines differ by two.

row	sub-claim	refusal today	sources read	base result and steps, or the objection	verdict	owning row
F8	(a) a swap of the two arms of a declared pairwise contrast, in the latent tail	`_MULTI_ARM_REFUSAL`	DoWhy `2116d5c`, `_include_confounders_effect`, the binary flip branch	the binary flip applied to the two arms only, \(g'_a = (1-s)g_a + s g_b\). The latent is independent of the data, so the law is a declaration. The swap is label-invariant by construction	moves: natural extension	X47 (a)
F8	(b) a declared \(K \times K\) transition	`_MULTI_ARM_REFUSAL`	Hu et al. (2022), arXiv:2012.06093, Remark 2	the law is well defined, but no source states the reading of its strengths	not adopted	none
F9	(a) a cluster-level latent; (b) a row-level latent on a clustered fit; (c) a mixed latent with a declared \(\rho\)	`_CLUSTERED_REFUSAL`	DoWhy `2116d5c`, `_include_confounders_effect`, the row-level draw; `treatSens` 3.0, `R/genU_contY.R` and `R/treatSensMLM.R`	the caller declares `latent_level`. Each level is a declared law. `treatSens` is a design precedent for the level rule, not a base. A row-level latent is refused when the treatment is cluster-constant	moves: natural extension	X47 (b)
F10	(a) a block flip calibration; (b) a block \(R \cdot \mathrm{sd}\) calibration of one logical categorical covariate	the categorical calibration refusal in `_simulated_confounding_request.py`	DoWhy `2116d5c`, `preprocess_observed_common_causes` and `_infer_default_kappa_t`	calibration over every encoded column of one covariate, fitted unpenalized. An unpenalized fit makes the strength invariant to the reference level. The shipped helper uses a default L2 penalty, so the contract changes the fit	moves: natural extension	X47 (c)
F11	hold the supplied weights fixed on an estimated-weight fit	`_ESTIMATED_WEIGHT_REFUSAL`	`src/cleverly/data/causal_data.py`, the weight declaration; `drtmle.py`, the estimated-weight note	probability weights define the target population. Each cell targets the fit’s weighted population, \(E[w\,h(W, A', Y')]\). No step needs new theory	moves: no new theory	X47 (d)
F11	regenerate the weights on the perturbed data	`_ESTIMATED_WEIGHT_REFUSAL`	`treatSens` 3.0, `man/treatSens.Rd`	regenerated weights define a population that the fit never targeted. The `treatSens` weights are treatment-model weights, which the package refits as \(g\)	not adopted	none
F11	a latent inside the weight model	`_ESTIMATED_WEIGHT_REFUSAL`	Hartman and Huang (2024), abstract only	the refusal does not stop this request	not a claim of this row	none
F12	(a) outcome-axis cells; (b) a declared response mechanism that reads the covariates only	`_MISSING_OUTCOME_REFUSAL`, `_MISSING_TREATMENT_REFUSAL`	DoWhy `2116d5c`, `_include_confounders_effect`	\(\Delta \perp (A, Y) \mid W\) and an independent latent give \(\Delta \perp (A', Y') \mid W\), so missing at random holds given \((A', W)\)	moves: natural extension	X47 (e)
F12	(c, d) a response or observation mechanism that reads the treatment	as above	Chen and Heitjan (2023), abstract only	missing at random given \((A', W)\) fails under the flip, so the stored estimator does not identify the perturbed law. A refit with the original \(A\) as an auxiliary variable identifies it, but it is a different estimator	keeps: identification	F12
F13	(a) one time-fixed latent, uncensored terminal outcome, no time-varying covariate after the first treatment node	`_LONGITUDINAL_REFUSAL`	DoWhy `2116d5c`, `_include_confounders_effect`, per node	the DoWhy law at each treatment node with one shared latent. The narrowing removes covariates generated under the original treatment. A later treatment node whose mechanism reads an earlier treatment has the same structure, because its induced association with the latent includes the path through the earlier misclassified treatment. X48 must define that induced association or restrict to mechanisms that do not read an earlier treatment	moves: natural extension	X48
F13	(b) censoring or a survival outcome; a time-varying covariate after the first treatment node; a later treatment node whose mechanism reads an earlier treatment	`_LONGITUDINAL_REFUSAL`	Tan (2025), arXiv:2308.15725v1, Section 3.1	the flip breaks coarsening at random at each censoring node. A covariate generated under \(A_t\) depends on the latent given \(A'_t\). A later treatment node whose mechanism reads an earlier treatment stays here until X48 defines its induced association or restricts the mechanisms	keeps: identification	F13
F13	(c) a time-varying latent with treatment feedback	`_LONGITUDINAL_REFUSAL`	Tan (2025), arXiv:2308.15725v1, Equations (11) and (34)	Tan gives bounds and no law whose strengths name the feedback	keeps: identification	F13
F15	(a) the treatment axis on a fit with an intermediate	`_INTERMEDIATE_REFUSAL`	`mediationsens` 0.0.3, `sens()`	a flip of \(A\) makes the latent confound \(Z\) and \(Y\) through the misclassified \(A\). That path has no named strength	keeps: identification	F15
F15	(b) outcome-axis cells; (c) an intermediate-axis law that flips or shifts \(Z\) and shifts \(Y\) with \(A\) held	`_INTERMEDIATE_REFUSAL`	DoWhy `2116d5c`, `_include_confounders_effect`	the DoWhy law at the \(Z\) node, \(g_{Z'} = (1-s)g_Z + s(1-g_Z)\). The latent confounds \(Z'\) and \(Y'\) only, with named strengths	moves: natural extension	X47 (f)
F23	an outcome law on the logit scale of the declared `q_bounds`	a per-cell failure (`src/cleverly/utils/bounds.py`)	Zheng, D’Amour and Franks, arXiv:2102.09412, Section 4, Equations (12) to (15); DoWhy `2116d5c`, the linear outcome branch	a shift on the logit scale of \([a, b]\) stays in the support and is the identity at zero strength. A normal-score option has lower confidence	moves: natural extension	X47 (g)

Sensitivity bounds#

row	sub-claim	refusal today	sources read	base result and steps, or the objection	verdict	owning row
F25	(a) the risk-ratio E-value of a controlled direct effect, point and limit	`_DIRECT_EFFECT_REFUSAL`	Ding and VanderWeele (2016), arXiv:1507.03984v1, Proposition A.2; Smith, Mathur and VanderWeele (2021), arXiv:2005.02908v1, Result 1	the bias of the ratio splits into one factor per arm. That split is an identity. Proposition A.2 bounds each factor. Result 1, whose \(BF_s\) is a per-arm product, is the structural precedent. With equal strengths the inversion is the ordinary E-value of \(\sqrt{RR}\)	moves: natural extension	X49
F25	(b) the odds-ratio and Gaussian conversions on such a fit	`_DIRECT_EFFECT_REFUSAL`	as (a)	part (a) and the shipped conversions. Each keeps its approximation label	moves: natural extension	X49
F26	(a) confidence limits of the plug-in \(E_n[\hat\alpha^2]\) for a learned mechanism	`_PLUGIN_LIMITS_REFUSAL`	Chernozhukov et al. (2018), arXiv:1608.00060v7, Section 1.1; DoubleML `DoubleML._validate_sensitivity_elements`	\(E[\hat\alpha^2] - \nu_0^2\) has a term linear in the mechanism error. Section 1.1 records that bias for a non-orthogonal plug-in	keeps: regularity	F26
F26	(b) the plug-in limits with a declared known mechanism	`_PLUGIN_LIMITS_REFUSAL`	Chernozhukov, Cinelli, Newey, Sharma and Syrgkanis, arXiv:2112.13398v5, Theorem 4	\(E_n[\alpha(g_0)^2]\) is a sample mean of a known function: linearity, and the delta method for the ATT and ATC share	moves: natural extension	X26
F26	(c) the orthogonal \(\nu^2\) for any mechanism learner	none; `nu2_estimator="doubly_robust"` ships	arXiv:2112.13398v5, Section 4, Lemma 2 and Theorem 4. The repository’s earlier citation of version 6, not re-read here, numbers that lemma Lemma 3	the shipped estimator	no action	none
F16	(a) a sample estimate of the bounds of Tan (2025)	`_LONGITUDINAL_BOUND_REFUSAL`	Tan, arXiv:2308.15725v1, Section 3.1; Shen, Bacak and Kennedy, arXiv:2601.17779v1, Section 5.2	both sources record sample estimation as future work. Consistency of a nested program with generated pseudo-outcomes is no standard step	keeps: asymptotic	F16
F16	(b) an interval for that estimate	`_LONGITUDINAL_BOUND_REFUSAL`	Bonvini et al., arXiv:2210.04681v2, Section 6.1; Bruns-Smith and Zhou, arXiv:2302.00662v3	Section 6.1 leaves the time-varying constraint to future work. The bound functionals are not smooth	keeps: regularity	F16
F16	(c) an omitted-variable bound or an E-value on a longitudinal fit	`_LONGITUDINAL_BOUND_REFUSAL`; the longitudinal E-value refusal in `evalue.py`	Ding and VanderWeele (2016), Section 2; arXiv:2112.13398v5, Section 4	both bounds read a point exposure or a known regression. A sequential g-formula with treatment-affected covariates needs a per-period latent law	keeps: identification	F16

Collaborative and doubly robust compositions#

row	sub-claim	refusal today	sources read	base result and steps, or the objection	verdict	owning row
F18	(a) one declared partial candidate, parametric learners	`working_mechanism_plugin`	Saul and Hudgens (2020); Ju et al., arXiv:1806.06784v3	a stacked estimating equation over the outcome GLM \(\beta\), the candidate mechanism \(\gamma\), the fluctuation \(\epsilon\) and \(\psi\). The interval covers \(\psi_0\) when the outcome GLM is correct or the candidate mechanism equals \(g_0\)	moves: natural extension	X52 (a)
F18	(b) inference after the stopping-index selection	`working_mechanism_plugin`	van der Laan, Carone, Luedtke and van der Laan, arXiv:2307.12544v2, Section 2.2	a data-adaptive selection step. The adaptive framework needs a fixed oracle submodel	keeps: adaptive-optimum inference	F18
F19	(a) the per-arm design, parametric outcome, in sample; (d) its composite cells	`generated_design_plugin`	Benkeser, Cai and van der Laan, arXiv:1901.05056v1, Appendix F	a fully parametric stack, as for F18 (a). The design learner’s own score enters the stack, so condition (iii) is not needed	moves: natural extension	X58 for part (a); X52 (b) for part (d)
F19	(c) the shared design with a parametric outcome and design	`generated_design_plugin`	arXiv:1901.05056v1, Theorem 1	one finite-dimensional M-estimator on the \(K\) prediction columns, as for F18 (a)	moves: natural extension	X52 (c)
F19	(b) a flexible outcome learner	`generated_design_plugin`	arXiv:1901.05056v1, Sections 4.1 and 6	the term’s Riesz representer reads \(1/g_0\). No standard step carries it	keeps: asymptotic	F19
F19	(c) a nonparametric shared multinomial design	`generated_design_plugin`	arXiv:1901.05056v1, Theorem 1	Theorem 1 is scalar, and no result gives the vector expansion	keeps: regularity	F19
F19	(e) the cross-fitted construction; (f) local uniformity over the nonparametric model	`generated_design_plugin`	arXiv:1901.05056v1, Appendix D and Section 6	Appendix D states no theorem. Section 6 calls the estimator irregular	keeps: asymptotic	F19
F22	(a) a selector point with `id=`; (b) an outcome-adaptive point with `id=`	“C-TMLE has no clustered result” (`ctmle.py`)	van der Laan and Gruber (2009), WP 246, Section 4	a cluster as the unit with whole-cluster folds. Each point inherits the fixed-limit assumption for the selected mechanism	moves: natural extension	X51 (e)
F22	(c) a clustered interval	as above	F18, F19	the clustered versions of the X52 cells move with X52. Every other clustered interval inherits F18 (b) and F19	keeps: asymptotic	F22
F5	C1 C-TMLE ATT and ATC	“CTMLE does not support estimand(s)” (`ctmle.py`)	van der Laan and Gruber (2009), WP 246, Theorem 2	the clever covariate and the treated covariate law read \(g\), so the target is not a function of \(Q\) alone	keeps: regularity	F5
F5	C2 C-TMLE natural course, PAR and PAF with missing outcomes, and cross-fitted arm-indexed missing outcomes (points)	the natural-course refusal in `tmle.py`	WP 246, Section 2.4	the missingness model stays full-covariate and off the selection path, as the shipped selector fits it. The natural-course estimate is then an ordinary TMLE stacked beside a C-TMLE arm mean	moves: natural extension	X51 (a)
F5	C3 C-TMLE regimes (points)	`_REGIME_REFUSAL`	WP 246, Theorem 2	the target reads \(Q\) and the covariate law only: an indicator reduction to \(1\{A = d(W)\}\)	moves: natural extension	X51 (b)
F5	C4 C-TMLE incremental; C5 modified treatment policies	“CTMLE and incremental= are not combined”; `_POLICIES_REFUSAL`	WP 246, Theorem 2	each candidate \(g\) defines a different parameter, so collaborative consistency fails	keeps: regularity	F5
F5	C6 C-TMLE discrete-arm MSMs (points)	`_MSM_REFUSAL`	WP 246, Theorem 2	a fixed full-rank projection of the arm means, with a declared selection criterion	moves: natural extension	X51 (c)
F5	C7 C-TMLE controlled direct effect	“CTMLE does not compose either collaborative strategy with an intermediate outcome”	Schnitzer, Sango, Ferreira Guerra and van der Laan (2020), abstract only	a path over two mechanisms is no standard step. The abstract suggests a published point method, and its first-hand read is pending	keeps: regularity, first-hand read pending	F5
F5	C8 C-TMLE selector with a missing treatment (point, in sample)	the composite refusal in `composite.py`	WP 246, Section 2.4	the composite indicator reduction, with both observation factors full-covariate and off the path	moves: natural extension	X51 (d)
F5	D1 DR-TMLE ATT and ATC; D3 modified treatment policies; D4 incremental; D6 controlled direct effect	“DRTMLE does not support estimand(s)”; the axis refusal in `drtmle.py`	Benkeser, Carone, van der Laan and Gilbert (2017), PMC5793673, Section 4 and Section 1; Díaz et al., arXiv:2006.01366v4, Theorem 2	Section 4 names other parameters as future work, and Section 1 names several time points. Theorem 2 assumes that the regime does not depend on \(P\)	keeps: asymptotic	F5
F5	D2 DR-TMLE regimes	the axis refusal in `drtmle.py`	PMC5793673, Theorem 1	an indicator reduction to \(1\{A = d(W)\}\). A known stochastic policy is a known tilt in \(W\) for each arm	moves: natural extension	X50 (a)
F5	D5 DR-TMLE discrete-arm MSMs	the axis refusal in `drtmle.py`	PMC5793673, Theorem 1	a fixed full-rank projection. Each tilt reads \(W\) only, so it leaves \(\bar Q_0\) and \(g_0\) unchanged	moves: natural extension	X50 (b)
F5	D7 DR-TMLE natural course, PAR and PAF with missing outcomes	the natural-course refusal in `tmle.py`	PMC5793673, Theorem 1	Theorem 1 with \(\Delta\) as the treatment and \((A, W)\) as the covariates, then a stack, linearity, and the delta method	moves: natural extension	X50 (c)
F5	D8 `evaluation=` on the composite	`COMPOSITE_NESTED_REFUSAL`	`docs/technical-reference/dr-tmle/index.md`, the evaluation companion	a remainder diagnostic, not an inference claim	moves: no new theory	X50 (d)
F5	D9 `reduced_crossfit="nested"` on the composite	unreachable: the cross-fit refusal comes first	`dr-tmle/targeting.md`, term [B]	the nested construction is the candidate route for the open term [B]	keeps: asymptotic	F21
F5	D10 the omitted-variable bound on DR-TMLE; D11 on C-TMLE	the bound refusals in `omitted_variable.py`	arXiv:2112.13398v5, Section 4	the shipped plug-in bound recentred on the point, under a consistent mechanism. The owner chooses the scope	moves: no new theory	X50 (e); X51 (f)
F5	D12 DR-TMLE estimated weights with a guard	`estimated_weight_plugin`	`src/cleverly/data/weighting.py`, the conditional reading	the conditional reading of the weights does not use efficiency, so it covers the fixed-weight DR-TMLE curve	moves: natural extension	X50 (f)

Cross-fitting, bootstrap and clusters#

row	sub-claim	refusal today	sources read	base result and steps, or the objection	verdict	owning row
F21	(a) fold-specific targeting with missing outcomes	“no direct interval result covers fold-specific targeting” (`tmle.py`)	Montoya et al. (2023), Section 3.2, Step 2(c)	the composite indicator reduction. The coefficient is fit on the validation rows, which is not the failing training-row update	moves: natural extension	X53 (a)
F21	(b) the median over a fixed number of splits	“no audited result covers the median over repeated splits” (`tmle.py`)	the shipped stacked contract: Zheng and van der Laan (2011), WP 273, Theorem 1, with the composite indicator reduction, for a fixed number of splits. Corollary 3.3 of Chernozhukov et al. (2018) was not re-read, and the argument does not need it	each draw is \(\psi_0 + P_n D + o_P(n^{-1/2})\) with common limits, so the median is too. No independence between draws is needed. The split-dispersion term is \(o_P(n^{-1})\)	moves: natural extension	X53 (b)
F21	(c) a cross-fitted `TMLE` with a missing treatment	the composite cross-fit refusal in `composite.py`	the shipped stacked contract; the composite verdict (a) of the natural-extension verdicts	a product of fold-local factors estimates \(P(C_a = 1 \mid W)\). The indicator reduction carries it into the stacked contract	moves: natural extension	X53 (c)
F21	(d) a cross-fitted `DRTMLE` with `delta=` or a missing treatment	the cross-fit refusals in `drtmle.py` and `composite.py`	`dr-tmle/targeting.md`, terms [A], [B] and (E)	the package’s own argument leaves term [B] open, and no cross-fitting lemma reaches it. The nested construction is the candidate route	keeps: asymptotic	F21
F21	(e) shift, regime, MSM and controlled-direct-effect targets, cross-fitted with missing outcomes	`_refuse_cross_fitted_missing_off_contract` (`tmle.py`)	Díaz et al., arXiv:2006.01366v4, Section 3, Section 5.2 and Theorem 3	Theorem 3 with \((A, \Delta)\) as the exposure. An MSM adds the projection	moves: natural extension	X53 (d)
F21	(e) incremental and risk-ratio tilt targets, cross-fitted with missing outcomes	as above	arXiv:2006.01366v4, Theorem 2	Theorem 2 assumes that the regime does not depend on \(P\). Kennedy (2019) and Kim et al. (2021) are the bases, and the review did not re-read them	moves: pending source read	X53 (d)
F21	(f) the four stacked follow-ups	each named in the stacked contracts	Zheng and van der Laan (2011), Section 2	each is outside the hard stop and needs its own contract	moves: natural extension	X53 (e)
F2	(a) full-refit bootstrap consistency with data-adaptive or cross-fitted nuisances	a cross-fitted point bootstrap, and every `DRTMLE` and `CTMLE` bootstrap, prints as a diagnostic. An in-sample `TMLE` fit with a data-adaptive learner still publishes the interval	Tang and Westling, arXiv:2404.03064v2, Theorem 3.4 and Section 2.3	condition (B2) is unverified for general learners. The paper does not treat cross-fitting, and a resample can put copies of one row in both folds. The registered study measured only correctly specified nuisances, so this row holds the data-adaptive case	keeps: asymptotic	F2
F2	(b) the natural-course, PAR and PAF bootstrap with missing outcomes, in sample, finite-dimensional nuisances	“set n_bootstrap=0” (`tmle.py`)	arXiv:2404.03064v2, Theorem 3.4	the bootstrap of a smooth M-estimator, and a registered study	moves: no new theory (Tang and Westling, Theorem 3.4), with a registered study	X55 (a)
F2	(c) the HAL-TMLE bootstrap	no request reaches it; no HAL learner ships	Cai and van der Laan (2020), arXiv:1905.10299, Section 3.6, Theorem 4	in sample, with the HAL bound \(C\) fixed at the selected value and a per-replicate update	moves: published support	X55 (b)
F2	(c) the targeted bootstrap of Coyle and van der Laan	no request reaches it	Coyle (2017), Berkeley dissertation (eScholarship 9kh0b9vm), Section 1.4 and Appendix A, Theorem 9 (thesis numbering); the book chapter of Coyle and van der Laan (2018) was not read	a model-based bootstrap from the targeted fit, with Super Learner and TMLE refitted and retargeted on each sample. Theorem 9 needs the efficient score solved to the order it states, a negligible remainder, a uniform Donsker class and \(L_2\) convergence of \(D^*\). The local read records that order as \(o_P(1/n)\), and X55 confirms it before it ships	moves: published support (dissertation)	X55 (c)
F2	a row-resampling (nonparametric) targeted bootstrap for learners other than HAL	no request reaches it	arXiv:1905.10299, Theorem 4; Coyle (2017), Theorem 9	Theorem 4 covers a HAL learner only. Theorem 9 covers a model-based bootstrap under a uniform Donsker class, which X55 (c) holds	keeps: asymptotic	F2
F28	items 1 to 8: intervals at 4 to 9 clusters (4 to 19 for `LTMLE`), a fold-evaluated degrees-of-freedom rule, another small-sample reference, the fold-evaluated variance and point at few clusters per fold, a cross-fitted `DRTMLE` cluster variance, complex learners at about 20 clusters	`few_cluster_plugin`, and the red cells that `F28` owns	Nugent et al., arXiv:2208.09508, Section 2.2; Benitez et al., arXiv:2110.09633v3, Sections 3.1.2 and 3.2.1; Fang et al., arXiv:2510.08438, Section 3.3; Wang, Park, Small and Li (2024), accepted manuscript PMC11795269, Section 4.2, Remark 3	the theory of a cluster as the unit is the same at 4 clusters as at 10. The shipped \(t_{J-2}\) is a recommendation with no derivation. Each item needs a registered measurement, not a theorem. Remark 3 of the accepted manuscript recommends parsimonious models at about 20 clusters, a finite-sample warning. arXiv:2210.07324v3 gives no counts	moves: no new theory	X54 (a) to (h)
F28	an interval below 4 clusters	`few_cluster_plugin`	the Eligibility failing example	below 4 clusters the reference has fewer than 2 degrees of freedom. At 2 clusters or fewer no independent-cluster argument applies	keeps: arbitrary dependence	F28
F31	an interval at a boundary node, at a fixed law	`constant_node_plugin`	none beyond the package	a zero-event node has exponentially small probability at a law with interior hazards and a node risk set that grows linearly in \(n\). An interval that differs from the ordinary one only on that event is valid at every fixed law under the shipped theorem. A law with a node hazard of exactly 0 or 1 needs a conservative construction, or F31 keeps it	moves: no new theory	X54 (i)
F31	the rare-event local regime, \(h_n \to 0\)	`constant_node_plugin`	Andrews (1997), abstract only	boundary theory assumes the truth on the boundary. No read source treats a hazard that shrinks with \(n\)	keeps: asymptotic	F31
F7	(a, b) blocked-temporal and rolling-origin folds for a dependent calendar series	iid and grouped folds only	Ballinari and Wehrli, arXiv:2411.10009v2, Theorem 4; Semenova et al., arXiv:1712.09988v6	both results are DML under mixing, not TMLE. No standard step carries mixing	keeps: serial dependence	F7

Estimands and compositions#

row	sub-claim	refusal today	sources read	base result and steps, or the objection	verdict	owning row
F3	(a) the risk under eliminated competing events	no request reaches it	Young, Stensrud, Tchetgen Tchetgen and Hernán, arXiv:1806.06136v3, Sections 5, 6.1 (Equation 23) and 8.1	the competing event is a censoring event. The shipped censored `LTMLE` fits a composite censoring indicator \(1\{C_t = 1, D_t = 0\}\) placed after \(D_t\)	moves: published support	X56 (a)
F3	(b) separable effects with baseline covariates	no request reaches it	Stensrud et al. (2022), arXiv:1901.09472v2, Section 5.2, formula (5); Wang et al., arXiv:2304.04904v1	the Wang et al. estimator with \(Z_t := D_t\)	moves: source audit	X56 (b)
F3	(c) separable effects with time-varying common causes	no request reaches it	Stensrud, Hernán, Tchetgen Tchetgen, Robins, Didelez and Young (2021), Lifetime Data Analysis 27(4):588-631, Section 6.2, Equation (19), which is Section 7.2 of arXiv:2004.14824v2; Wang et al., arXiv:2304.04904v1	the Wang et al. estimator with \(Z_t := (L_{A_D,t}, D_t)\). No source records a defect	moves: source audit	X56 (c)
F3	(d) other longitudinal estimands	none	none	the sentence names no target	not a claim of this row	none
F6	(a) the MNAR tilt on a shift fit; (a’) on an incremental fit	`_refuse_continuous`, `_refuse_incremental` (`missingness.py`)	the shipped tilt law; Kennedy, arXiv:1704.00211v3	the point curve \(E[\bar Q_\gamma(d(A, W), W)]\). At \(\gamma \neq 0\) the spread columns use the plug-in names and claim no interval	moves: natural extension	X57 (d)
F6	(b) a controlled direct effect under a tilt	“incremental= and intermediate= are not combined” (`tmle.py`)	Kennedy, arXiv:1704.00211v3, Lemmas 1 to 3; part (s) of the natural-extension verdicts	a two-node `IPSI` plan, a tilt node and then a label node \(Z = z\), which ships	moves: established method (a redirection)	X57 (e)
F6	mediation under a tilt	none; no request reaches it	Hejazi et al., arXiv:2009.06203v4, Lemma 3 and Theorems 3 and 4	the incremental mediation effects of that paper	moves: published support	X5
F17	(a, c) two intervention axes in one fit	the axis refusals in `tmle.py` and `interventions/base.py`	none beyond the standard steps	separately targeted fits on the same rows, then a fixed-dimension stack. An incremental axis does not share \(g^*\) with a regime axis	moves: natural extension	X57 (a)
F17	(b) an MSM over declared regimes, shifts or tilts; (b’) an arm MSM beside another axis	the MSM refusal in `tmle.py`	`msm-projections.md`, the MSM score	the shipped MSM score with the arm indicator replaced by \(1\{A = d(W)\}\) or \(q(a \mid W)\). The score carries effect modifiers \(V\). Part (b’) is a stack	moves: natural extension	X57 (b), (c)
F17	(d) a learned rule beside another axis	the learned-rule axis refusal in `learned.py`	van der Laan and Luedtke (2015), PMC4517487, Theorem 6	under C3 the CV-TMLE has a fixed limit curve, so the fixed-dimension stack applies, as for the contrast of part (c)	moves: source audit, with part (c)	X11 (c)
F27	1 an interval at an exceptional law	none; the boundary study measures it	PMC4517487, Theorem 6; Hubbard et al. (2016), Theorem 1; Luedtke and van der Laan, arXiv:1603.07573, Sections 4.2 and 5; Fava, arXiv:2610.02944v1	both theorems need a limit curve. Fava covers a known-propensity IPW score only	keeps: adaptive-optimum inference	F27
F27	2 the `weak_blip` law	none; the boundary study measures it	PMC4517487, Theorem 6	the law is within C3. The red cell is a finite-sample limit	keeps: an evidence record	F27
F27	3 `repeats` above 1, with the mean	the learned-rule scheme refusal	PMC4517487, Theorem 6	under C3 in its fixed-limit form each split is asymptotically linear with a common limit, so the mean follows by linearity	moves: natural extension	X11 (i)
F27	4 the full-refit bootstrap	the learned-rule scheme refusal	none supplies the argument	each resample relearns the rules, so the bootstrap law of a data-adaptive target is not given	keeps: asymptotic	F27
F27	5 a continuous treatment	the learned-rule data refusal	the Eligibility failing example	a fixed dose rule is not pathwise differentiable	keeps: regularity	F27
F27	6 `intermediate=`	the learned-rule data refusal	PMC4517487, Section 2 and Theorem 6	a learned rule at \(A\) and the static \(Z = z\) form a two-time-point rule pair	moves: source audit	X11 (j)
F27	7 `weights=`; 8 `id=`; 9 `strata=`	the learned-rule data refusal	PMC4517487, Theorem 6	the chain rule with a known weight; a cluster as the unit; a finite partition after X28	moves: natural extension	X11 (k), (l), (m)
F27	10b `CTMLE` with `learned_rule=`	the learned-rule configuration refusal	F18, F19	the selector and outcome-adaptive paths have no interval for any target	keeps: adaptive-optimum inference	F27
F27	a learned rule with missing outcomes (`delta=`); the missingness tilt on a learned-rule fit	the learned-rule data refusal; `LEARNED_RULE_TILT_REFUSAL`	PMC4517487, Section 1	van der Laan and Luedtke (2015) treat complete outcomes, and no read source gives a cross-fitted learned-rule fit with missing outcomes. The earlier citation of F21 named a different row	keeps: asymptotic	F27
F27	13 a learned rule with a missing treatment	`_LEARNED_RULE_REASON` (`composite.py`)	PMC4517487, Section 1; the composite contract	the value is cross-fitted. No read source gives a cross-fitted learned-rule fit with a missing treatment, as for missing outcomes	keeps: asymptotic	F27
F27	10a a fold-evaluated `DRTMLE` learned rule	the learned-rule configuration refusal	none; the `DRTMLE` refusal of `cv_evaluation=True`	no read source gives a fold-evaluated DR-TMLE	keeps: asymptotic	F27
F27	11a the omitted-variable bound; 11b simulated confounding and `refute`	`LEARNED_RULE_BOUND_REFUSAL`, `LEARNED_RULE_REPLAY_REFUSAL`, `LEARNED_RULE_REFIT_REFUSAL`	PMC4517487, Theorem 6; `_simulated_confounding_fixed.py`, the regime replay	linearity over frozen fold rules after X9. The replay needs a new fold-evaluated path, because the shipped replay re-evaluates declared regimes only	moves: natural extension	X11 (n), (o)
F27	11c an E-value for a learned-rule contrast	`LEARNED_RULE_EVALUE_REFUSAL`	`evalue.py`, the arm-contrast check	no E-value of a regime contrast ships. A regime contrast is not a contrast of two exposure levels	keeps: regularity	F27

The conditions each moved part inherits#

owning row and part	inherited conditions
X47 (a) to (g)	each law is a declaration with a stated reading, not an estimator claim. No interval claim. The fit’s own identification conditions hold for the unperturbed data
X47 (d)	the supplied weights stay fixed. Each cell targets the fit’s weighted population
X47 (e)	a declared response mechanism that reads the covariates only, \(\Delta \perp (A', Y') \mid W\)
X48	one time-fixed latent shared by every treatment node; uncensored terminal outcome; no time-varying covariate after the first treatment node; a defined induced association at a later treatment node whose mechanism reads an earlier treatment, or no such node
X49	\(Y_{az} \perp A \mid W, U\) and \(Y_{az} \perp Z \mid A, W, U\), with \(U\) a baseline latent that \(A\) does not affect; positivity given \((W, U)\). Strengths are suprema over \(w\). The odds-ratio and Gaussian conversions keep their approximation label
X50 (a) to (f)	Theorem 1 of Benkeser et al. (2017), in sample: its rate conditions and its positivity bound. Part (e) holds under a consistent mechanism
X51 (a) to (f)	points only. Every fit keeps `working_mechanism_plugin` or `generated_design_plugin`. Observation factors stay full-covariate and off the selection path. Part (e) inherits the fixed-limit assumption of WP 246, Section 4
X52 (a) to (d)	unpenalized parametric learners read from the fit; inactive `g_bounds` at the limit; coverage of \(\psi_0\) needs the outcome GLM correct, or, for part (a) only, the candidate mechanism equal to \(g_0\). Otherwise the interval covers the pseudo-true limit. Regular in the declared outcome model, with no claim over the nonparametric model. Part (d) uses the few-cluster rules
X53 (a) to (e)	the stacked missing-outcome contract: package-generated folds, missing at random given \((A, W)\), positivity of the composite mechanism, and the rate conditions of the base theorem
X54 (a) to (i)	the shipped cluster-as-unit theory. Part (i) holds at a fixed law with interior hazards and a node risk set that grows linearly in \(n\)
X55 (a) to (c)	in sample. Part (a) needs finite-dimensional nuisances, read off the fit’s own learners. Part (b) needs the HAL learner with \(C\) fixed at the selected value. Part (c) needs the efficient score solved to the order of Theorem 9, a negligible remainder, a uniform Donsker class, \(L_2\) convergence of \(D^*\), and the model-based bootstrap from the targeted fit. Its rates \(o_P(n^{-1/4})\) under the targeted fit, positivity and bounded sectional variation are shown for the treatment-specific mean only (Lemma 2, Corollary 1)
X56 (a) to (c)	the node order of the competing event and the censoring indicator. Parts (b) and (c) need the user-declared partition of \(L_t\), whose dismissible-component conditions no data check can verify
X57 (a) to (c)	each axis keeps its own identification and positivity conditions. Each composition reduces exactly to each standalone fit
X57 (d)	no interval at \(\gamma \neq 0\)
X11 (i) to (o)	condition C3 in its fixed-limit form (van der Laan and Luedtke 2015, Theorem 6). Part (l) uses the few-cluster rules. Part (m) waits on X28 and part (n) on X9
X26	a declared known mechanism, with no truncation of it
X58 (a) to (d)	a declared regular parametric outcome model; a declared smooth parametric adaptive mechanism; the stated `g_bounds` behavior. Coverage of \(\psi_0\) needs a correct outcome model, or a clipped mechanism limit equal to \(g_0\). Otherwise the interval covers its limit \(\psi^*\). Pointwise, and regular in the declared model. Part (a) shipped on the joint-equation route, and the full-mechanism route is not built

Abstract-only sources#

The review read the sources below from their abstracts only. Each carries no verdict weight. A row that cites one keeps the readiness “pending source read” or records the source as an open read.

source	where it bears
Ou, Tang and Chang (2023), arXiv:2301.12396v1	F9, clustered sensitivity analysis
Hartman and Huang (2024), arXiv:2206.07119v2	F11, omitted variables in survey weights
Strezhnev (2026), arXiv:2607.13334v1	F9, clustered treatment assignment
Chen and Heitjan (2023), arXiv:2302.03740v1	F12, missingness from unmeasured confounding
Schnitzer, Sango, Ferreira Guerra and van der Laan (2020), Biometrics 76(1):145-157	F5 C7, a collaborative controlled direct effect. Its first-hand read is pending
arXiv:2411.02771v3	F7, a search lead for dependent data
Andrews (1997)	F31, estimation on a boundary
van der Laan, Benkeser and Cai, arXiv:1908.05607v2	F26, an undersmoothed HAL plug-in
Smith and VanderWeele (2019)	F25, natural-effect sensitivity
Sjölander and Waernbaum (2024)	F25, bounds for a direct effect
Stensrud et al., arXiv:2004.14824v1	F3 (c). The verdict rests on the full text of version 2 and on the published version of 2021

The review did not read two sources. The access to Carnegie, Harada and Hill (2016) failed, so the review cites only its implementation, treatSens 3.0. Scharfstein, Rotnitzky and Robins (1999) was not read, and no verdict cites it.

VanderWeele (2010), Epidemiology 21(4):540-551, PMC4231822, Theorem 1, is context for X49 and not its base. That theorem assumes that the latent confounds only the intermediate and the outcome. The bias split of X49 is an identity that also lets the latent confound the treatment.