Extended Data 4; SI: acquisition
Randomised evidence depth
The initial controlled language gives a 4.3-fold shift over the analysed quantiles. The crossed study finds positive delays in all six architecture–construction cells, with all control intervals spanning zero.
Deeper evidence makes learning arrive later.
Evidence depth is assigned at random across three architectures and two tasks. Compare the resulting delay with fixed-depth and shuffled-label controls.
Loading effects…
The initial depth experiment
Loading recorded measurements…
How much does deeper evidence delay learning?
Loading recorded comparison…
Methods and interpretation
The initial display covers the 10th–20th acquisition quantiles. The crossed study uses eight seeds per cell and matched blocks. Persistent preference and early unsigned movement are different outcomes.
Procedure in the paper: Generate the paired controlled languages, randomise evidence assignments, train from fixed paired seeds and batch streams, and apply the registered acquisition threshold, controls and crossed bootstrap.
Code and data
v0.1.0 · 93 KB · View source on GitHub ↗
The full randomized-depth experiment with generated data, controlled training, and descriptive additive fits across depths and margin bins.
v0.1.0 · 805 KB · View source on GitHub ↗
Exact local generation, calibration, training and causal/mechanism analysis for three architectures crossed with two controlled evidence constructions.
Each standalone package includes code, shared helpers, required small inputs and reference results, with setup and commands in its README. Model weights and public datasets are obtained separately where needed.
Download complete source (v0.1.0) for all experiments and the companion website.