Interventional-twin audits for post-probe decision routing under hidden-factor shift

Artikel i vetenskaplig tidskrift, 2026

Interactive systems can encounter states that look equivalent before a low-energy probe even though a hidden mass produces different motion and therefore calls for a different commit route. Aggregate success cannot reveal whether a router used this observed post-probe effect or a correlated pre-probe shortcut. We introduce an interventional-twin audit for this question. The audit predeclares observable channels, constructs pairs that match before probing but diverge in observed effects, freezes a non-oracle input schema, and evaluates cue reversal, directional effect swaps, leakage, route failure, and selective re-probing. Across 180 family–seed units in six fixed synthetic parameterizations, direct post-probe routing attains 0.871 reversed-cue success versus 0.500 without the effect channel. When the factual effect is replaced by its matched counterpart while visible history and probe action are fixed, the learned router attains 0.759 bilateral directional consistency versus 0.000 without probe-effect access. A CausalWorld Pushing construction provides public evidence: matched state and frozen-RGB cases establish post-probe input value, and a separately frozen 20-run full-contract audit passes directional-permutation and grouped leakage gates. A 20-run matched-pair reliability diagnostic preserves pair identifiers while effect noise and missingness increase; the overall paired gain remains positive under severe corruption, while the severe worst-mass interval crosses zero. An architecture–capacity–budget stress shows positive post-probe gains but cell-dependent architecture ordering. A calibration-frozen risk–coverage sweep finds no safe positive-coverage point under three shortcut stresses. The result is a controlled, reproducible workflow for testing effect-responsive routing and reporting where evidence becomes inconclusive.