ADL — AI-Directed Learning

Which intervention is worth trying?

Can a system separate from the learner choose which learning intervention to try? ADL separates the learning model from intervention selection. Its question is whether quality improves at the same total budget, rather than merely whether fewer model calls are made.

The latest quality comparison

In the F2-041 report dated 13 October, the candidate evaluates two coordinate proposals and selects the one with lower training loss. The preregistered hypothesis requires at least three percentage points of terminal holdout accuracy improvement against each of four baselines.

The difference against the strong uniform baseline is about 0.088 percentage points. The upper uncertainty bound is below three points, so the primary hypothesis closes as FALSIFIED. Gains against weaker comparisons do not change that decision.

Why cost stays separate

Forward calls are about 36.4% lower, while intervention-controller cost is about 10.5% higher. An overall cost advantage is not established. Fewer model calls alone do not mean cheaper learning.

The next candidate

F2-043 considers a weight-norm penalty under a separate preregistration. Implementation is pending at the source cutoff; effectiveness has not been tested. F3 remains closed and Conflux replication unverified.

Source note

Sources reviewed on 14 October 2026: the ADL README, F2-041 two-proposal report and F2-043 preregistration protocol. Figures are reported evidence; no experiment or model call was run for this page. Failure of one candidate does not refute the whole ADL programme.