Action disagreement after changing candidate order.
Experiment E025 / terminal aggregate
The shuffle moved.
The signal didn't.
Candidate order changed the selected action often—but repeated calls with the same order were already unstable. After subtracting that baseline, the remaining order effect was small and uncertain.
48 positions192 callscorrected analysis v1.1
Baseline disagreement between identical-order calls.
A second identical-order baseline.
95% interval -2.1 pp to 13.5 pp; randomization p=0.191.
What survived
Instability is real.
Its cause is not settled.
The observed 42.7% cross-order disagreement is not itself a causal order effect. The model also disagreed 39.6% and 35.4% of the time under repeated identical orders. The adjusted estimate is 5.2 percentage points, with an interval spanning −2.1 to 13.5.
Evidence boundary. Action stability for one dated Luna alias cohort on development-exposed Black positions; not move quality, Elo, search, or a universal null.
Analysis correctionE025-analysis-amendment-1
one retryable undecodable response is no longer a terminal unknown failure. The corrected primary result is disclosed here rather than silently replacing the earlier analysis.
Limits carried into the page
- interval includes zero and positive effects up to about 13.5 percentage points
- two replicates per order
- development-exposed Black-only positions
- one dated Luna alias cohort without immutable provider snapshot ID
- provider backend heterogeneity
- no move-quality evaluator
- protocol-specific