What the validation programme established#
The active registered canonical DR-TMLE study
now adds a theory-first comparison with pinned R drtmle on the paper’s binary complete-data law.
Most paired cells establish bounded equivalence, but none establishes the prespecified
coverage-superiority route; one paired cell is inconclusive. The both-correct calibration and
root-n cells pass, while both one-correct robustness cells fail their finite-sample bias rule.
Those red results are committed evidence, not exceptions to the theorem: the theorem remains
conditional on rate and remainder premises that a fitted dataset cannot certify.
The older drift-law programme below asks a different, harder question. It is retained as historical evidence rather than silently generalized to the registered paper law.
This page’s claims rest on a closed programme of six pieces: a theoretical audit against the
sources, a targeting-and-exit study, a controlled coverage demonstration, a reference study for
the reduced regressions, a construction ablation, and a terminal experiment. The
evidence index records the acceptance instruments. The
programme itself, including its study harnesses, replicate records, differential diagnostics,
dispatch workflows, and working notes, is archived at the drtmle-validation-archive-2026-08 tag
rather than on main. In summary:
Established. The implementation is faithful to Theorem 1 in five respects.
The corrected curve is the Gateaux derivative of the parameter.
The sign of the mechanism correction is the appendices’ orientation.
The reported variance is Theorem 1’s.
The three score equations are solved at the state returned.
The interval beats a plain TMLE’s by a material margin in the drift cells, on that design. It does not show validity.
Not established in that archived drift-law programme, and recorded as such. Three things.
Nominal coverage anywhere in that study,
the best reading being 0.880. A localized cause for that shortfall: a six-contrast construction
ablation over 2,496 fits returned a null on its primary column, and a terminal experiment over
both a selection and an independent audit cohort nominated nothing. And any src/ change
justified against the theorem.
The reported se runs about 10% short of the spread it covers in one drift cell, and about 16%
long in the other. The programme attributed the shortfall to a second-order remainder that does
not vanish at these sizes. A committed probe, reviews/rm40/drift_attribution.py, tested that
reading under a rule fixed before it ran. The shortfall is not attributed. The conditioning term of the
inconsistent nuisance, evaluated at the mean fit, does not account for it.
cell |
predicted root-n bias |
archived root-n bias |
probe outcome |
|---|---|---|---|
|
-0.067 |
+3.33 |
refuted |
|
+0.021 |
+1.44 |
refuted |
The g-drift mechanism learner converges to an additive approximation of g_0, so that cell is
not inside Theorem 1’s union condition. The probe measures that error as 0.002 in the L2 norm.
The probe log also reports a sensitivity row that the rule does not score. It averages the term
over the fits, so it includes the random part of the mechanism fit. It predicts +1.70 for
g-drift, about half of the archived +3.33. The rule would read that row as inconclusive.
That is why the release claim is conditional validity, and why the conditions are stated on this page rather than assumed away.