NEWS

ria.test 0.2.1 (2026-07-21)

Add torch_seed to ria.test.control() to control Torch neural-network initialization and dropout independently of R's random-number generator.
Delay Torch device initialization until after R-randomized estimation steps, so first-use Torch initialization cannot change cross-fitting or Z-prime folds.

Restrict estimation to natural effects, randomized interventional analogues, and the TE - TE^R falsification test.
Replace effect with estimand = "test", "natural", or "ria"; the test is now the default.
Align returned estimand names with the paper and remove nuisance fits from the returned object.

Synced fold construction and influence-function identifiers with upstream fixes.