Published experiments validating the Quale evolution engine across multiple domains. Each domain demonstrates the engine's ability to produce emergent behaviour from topology evolution alone.
The foundational experiments that validated Quale. Agents evolved foraging, food discrimination, and social instincts in a grid-based survival simulation, with zero behavioural code.
NEAT-evolved connectomes discover foraging behaviour (eating and drinking) from survival pressure alone, without any behavioural code. Validated across 10 independent seeds with 787% average fitness improvement. Ablation study confirms behaviour persists when consumption rewards are completely removed.
Agents evolve 93% avoidance of dangerous food using sensory properties (colour, smell, texture) with no discrimination rules coded. A 7:1 safe-to-bad ratio emerges from sickness penalties alone. Fresh evolution outperforms seeded brains, demonstrating that prior knowledge can block new learning.
Two agents share an environment with dangerous food and peer state sensors. Discrimination comparable to Phase 2 is achieved, but cannot be attributed to social observation because item properties alone are sufficient. Establishes the multi-agent baseline and identifies the need to isolate the social signal.
Under complete sensory randomisation (no static cues to exploit), agents evolve to suppress eating near a peer and flee when the peer shows sickness. A 3.4:1 discrimination ratio emerges from social observation alone, with all 6 peer sensor connections evolving negative (avoidant) weights. Zero hidden nodes required.
The first application of Quale to a safety-critical domain. Three iterative experiments exploring signal compliance, speed governance, and emergent vigilance behaviour in evolved train driver connectomes.
First validation of Quale in a non-survival domain. Binary throttle mechanics and weak schedule penalties allowed evolution to discover that remaining stationary (71% idle) was a viable safety strategy. Identified three design flaws that informed the Rail-002 redesign. Confirmed domain-agnostic parser and engine work without code changes.
Continuous throttle, station stops, and terminus penalties produced a 27% fitness improvement over Rail-001. The evolved brain discovered proportional speed governance (current_speed to brake, +1.29), maximum-priority AWS acknowledgment (+2.0), schedule-pressure brake release (-1.80), and stress-induced acceleration. Attention remained disconnected because it had no causal effect on simulation outcomes.
Making signal perception dependent on the attention actuator produced a fundamentally different brain. Three sensors connected to attention (stress +1.70, speed_limit +1.26, brake +0.30), independently discovering stress-enhanced vigilance. The first functional hidden neuron computed a "station approach assessment" concept. The design principle validated: causal mechanisms produce realistic behaviour where direct rewards fail.
Dan Battye / 2026-03-14 / Experiment Rail-003b (parallel)
Parallel evaluation (20 cores, 70 seconds vs 35 minutes) produced a 7-hidden-neuron brain with 36 connections. Evolved a complacency countermeasure (familiar routes increase attention), dual-pathway signal processor distinguishing red from yellow, dead man's switch emergency brake, and fatigue-induced acceleration. Six documented human factors phenomena emerged from topology alone. The deepest and most behaviourally rich brain in any Quale experiment.