Experiments

V24: TD Value Learning

Period: 2026-02-19. Substrate: V22 + temporal difference value function (semi-gradient TD).

Hypothesis: Long-horizon prediction via value function $V(s) = \mathbb{E}[\sum \gamma^t r_t]$ integrates over all possible futures — inherently non-decomposable.

Metric	Seed 42	Seed 123	Seed 7	Mean
Mean robustness	1.034	0.998	1.003	1.012
Mean $\intinfo$	0.051	0.072	0.130	0.084
Final $\gamma$	0.748	0.746	0.741	0.745

Finding: Best robustness of any prediction experiment (1.012). Agents evolve moderate discount ( $\gamma \approx 0.75$ , horizon ~ 4 steps). But $\intinfo$ is seed-dependent. The bottleneck is architectural: a single linear value readout doesn't force non-decomposable structure.

V24 trajectories: robustness, integration, population, and TD error — **V24 evolution trajectories.** Top-left: robustness with dramatic spikes at drought boundaries (up to 1.5) — the highest transient robustness of any prediction experiment. Top-right: Φ shows the widest variance (0.02–0.25), with seed 7 reaching high values mid-evolution before declining. Bottom-left: population dynamics. Bottom-right: TD error decreases over evolution — value learning works, but doesn't force integration because the linear readout is decomposable.

Robustness and integration compared across V22, V23, V24 — **V22–V24 prediction experiment comparison.** Left: mean robustness. V24 (TD value) achieves the highest (~1.01), crossing the 1.0 threshold. V22 and V23 cluster below 1.0. Right: mean Φ. All three experiments overlap in the 0.06–0.10 range with high per-seed variance. Individual seed dots show no experiment consistently outperforms the others. The prediction target (scalar energy, multi-target, temporal value) does not reliably change integration — only architecture does (see V27).

Source code

v24_substrate.py — TD value function
v24_evolution.py — Evolution loop
v24_gpu_run.py — GPU runner