Route Familiarity as Emergent Complacency Modulator: Deep Topology Evolution Under Parallel Selection Pressure

Abstract

Rail-003b repeats the Rail-003 attention-gated perception experiment under parallel evaluation conditions, where each genome is tested across multiple route scenarios simultaneously. The parallel selection pressure drives significantly deeper topology evolution: the best genome develops 7 hidden neurons (5+ functionally active) and 36 connections, compared to Rail-003's single hidden neuron and 22 connections. The evolved connectome independently discovers a complacency countermeasure through route familiarity, where increasing familiarity with a route raises rather than lowers attention (weight +1.56 to attention). This contradicts the naive expectation that familiarity breeds complacency and instead mirrors professional driver training protocols that teach heightened vigilance on well-known routes. The topology also evolves a dead man's switch pattern for emergency braking, dual-pathway signal discrimination (separating red from yellow signal responses), and an arousal integrator hidden neuron that combines fatigue, stress, and braking state into a unified arousal signal. Parallel evaluation reduced wall-clock time from approximately 35 minutes to 70 seconds while producing a higher-fitness genome (99.69 versus 99.03), suggesting that parallel selection pressure favours the evolution of more robust, generalised strategies.

1. Introduction

1.1 Background

The Quale Rail series investigates whether NEAT-evolved connectomes can produce safe, realistic train driving behaviour from survival pressure alone. The series has progressed through increasingly complex environmental challenges:

• Rail-001: Evolved basic throttle and braking coordination from a distance-based fitness function.
• Rail-002: Evolved speed governance through penalty-based fitness, with agents learning to respect speed limits.
• Rail-003: Introduced attention-gated perception, producing context-dependent vigilance behaviour, a stress-to-attention pathway mirroring the Yerkes-Dodson law, and the first hidden neuron in the series.

Rail-003 demonstrated that realistic vigilance behaviour emerges when attention is causally necessary for accurate perception. The evolved topology used 22 connections, 1 hidden neuron, and 3 attention inputs to achieve a best fitness of 99.03. However, Rail-003 evaluated each genome on a single route scenario per generation, raising the question of whether the evolved behaviours would generalise across diverse operating conditions.

1.2 Motivation for Re-run

Rail-003b repeats the Rail-003 experiment with one critical change: parallel evaluation. Each genome is evaluated across multiple route scenarios simultaneously within each generation, rather than being tested on a single scenario. This modification serves two purposes. First, it tests whether parallel selection pressure produces qualitatively different topology from sequential evaluation. Second, it dramatically reduces wall-clock training time by exploiting multi-core parallelism, making the experiment practical for rapid iteration.

The hypothesis is that parallel evaluation will select for more generalised strategies, because genomes must perform well across diverse conditions rather than specialising for a single route configuration. This should produce deeper topologies (more hidden neurons, more connections) capable of context-sensitive responses to varying route characteristics.

1.3 Hypothesis

Parallel evaluation across multiple route scenarios will drive the evolution of deeper, more complex connectome topologies that generalise across operating conditions, producing higher fitness and more sophisticated emergent behaviours than the sequential evaluation used in Rail-003.

2. Materials and Methods

2.1 Configuration

The experimental configuration matches Rail-003 in all respects except evaluation mode. The following table summarises the key parameters.

Parameter	Value
Population size	150
Generations	500 (converged at 349)
Ticks per generation	5000
Track length	50 km (procedurally generated)
Stations	8 per track
Signal aspects	4 (green, double-yellow, yellow, red)
Speed limits	40, 60, 80, 110 km/h (zone-dependent)
AWS (Automatic Warning System)	Enabled; requires acknowledgement within 5 ticks
Attention gating	Multiplicative; degrades signal perception
Fatigue model	Linear accumulation, reset on station stop
Visibility conditions	Clear, fog, rain (randomised per generation)
Fitness function	Distance + station stops − signal violations − crashes − AWS failures
NEAT mutation rates	Default (add node 0.03, add connection 0.05)
Evaluation mode	Parallel (multiple route scenarios per genome per generation)

2.2 Reproducibility Note

All configuration parameters, random seeds, and route generation procedures are identical to Rail-003 except for the evaluation mode. The parallel evaluation uses the same fitness function aggregated across scenarios (mean fitness across parallel evaluations). This ensures that any differences in evolved topology or behaviour can be attributed to the parallel selection pressure rather than to changes in the fitness landscape.

3. Results

3.1 Fitness Progression

Table 1. Fitness, topology, and behavioural metrics across evolution.

Generation	Best Fitness	Avg Fitness	Species	Topology	Survival	Idle Rate
0	78.36	6.94	300	23n / 10c	100%	99%
5	81.57	32.38	57	23n / 8c	99%	93%
15	83.30	56.56	29	23n / 10c	99%	86%
50	95.85	78.52	11	25n / 19c	100%	63%
100	95.50	76.12	36	26n / 22c	99%	59%
150	97.07	79.00	3	28n / 29c	99%	59%
200	99.07	69.01	13	29n / 32c	96%	48%
250	99.03	71.55	9	30n / 36c	97%	48%
300	99.28	64.63	18	30n / 35c	95%	55%
349 (converged)	99.69

Fitness progression shows steady improvement from generation 0 (78.36) to convergence at generation 349 (99.69). The topology grew substantially over the course of evolution, from 23 neurons and 10 connections at generation 0 to 30 neurons and 36 connections by generation 250. Species count fluctuated considerably, dropping from 300 at initialisation to as few as 3 at generation 150 before rebounding. The idle rate, which measures the proportion of ticks where the agent takes no meaningful action, decreased from 99% to a minimum of 48%, indicating that the evolved agents became progressively more active and engaged with the driving task.

3.2 Comparison with Rail-003

Table 2. Rail-003b compared to Rail-003 (sequential evaluation).

Metric	Rail-003	Rail-003b
Best fitness	99.03	99.69
Convergence generation	240	349
Hidden neurons	2 (1 functional)	7 (5+ functional)
Connections	22	36
Minimum idle rate	71%	48%
Attention inputs	3	4
Throttle sensors	2	10
Wall-clock time	~35 min	70 sec

The comparison reveals that parallel evaluation produced a qualitatively different outcome. Rail-003b evolved 7 hidden neurons (with 5 or more functionally active) compared to Rail-003's 2 (only 1 functional). The connection count increased from 22 to 36, and the throttle output receives input from 10 sensors compared to Rail-003's 2. Despite requiring more generations to converge (349 versus 240), the wall-clock time dropped from approximately 35 minutes to 70 seconds due to parallelisation. The minimum idle rate fell dramatically from 71% to 48%, indicating that the parallel-evolved agents are far more active drivers.

3.3 Idle Rate Trajectory

Table 3. Idle rate progression across all four Rail experiments (at generations 0, 50, 200, and convergence).

Experiment	Gen 0	Gen 50	Gen 200	Converged
Rail-001	71%	72%	72%	72%
Rail-002	99%	72%	72%	71%
Rail-003	99%	71%	74%	74%
Rail-003b	99%	63%	48%	53%

The idle rate trajectory highlights a distinctive feature of Rail-003b. While Rail-001 through Rail-003 all converge to idle rates in the 71-74% range, Rail-003b drops to 48% at generation 200 before settling at 53% at convergence. This substantial reduction indicates that the parallel-evolved agents are using their actuators nearly twice as frequently as agents from previous experiments. The deeper topology enables more nuanced, continuous adjustments rather than the binary on/off behaviour seen in shallower networks.

4. Evolved Brain Topology Analysis

4.1 Overview

The best evolved genome from Rail-003b contains 30 neurons (18 sensor inputs, 5 actuator outputs, and 7 hidden neurons) connected by 36 enabled connections. This represents a significant increase in topological complexity over all previous Rail experiments. The following sections detail the connection weights for each actuator output and the functional roles of the hidden neurons.

4.2 Attention Neuron Connections

Table 4. Connections targeting the attention output (4 inputs plus bias).

Source	Weight	Interpretation
route_familiarity	+1.56	Familiar routes increase attention (complacency countermeasure)
speed_limit	+0.93	Higher speed limits raise alertness
current_speed	−1.07	Higher speed reduces available attention (cognitive load effect)
Hidden 568	+1.13	Familiarity-alert integration feeds into attention
Bias	+0.35	Baseline positive attention tendency

The most striking finding is the route familiarity connection (weight +1.56). Conventional wisdom suggests that familiarity breeds complacency: drivers who know a route well should pay less attention, not more. Yet the evolved connectome does the opposite. It increases attention on familiar routes. This is not a design artefact; it emerged because drivers who relaxed on familiar routes received degraded perception through the attention gate, made errors at predictable locations, and were removed from the population. The survivors were those whose genomes treated familiarity as a cue to heighten vigilance rather than relax it.

This mirrors professional driver training protocols in aviation and rail, where operators are explicitly taught to resist complacency on familiar routes. The NEAT process discovered this principle independently, through selection pressure alone.

Hidden 568 provides an additional familiarity-modulated input (+1.13), creating a two-stage processing pipeline where raw familiarity data is first integrated with alert state before feeding into the attention calculation.

4.3 Throttle Neuron Connections

Table 5. Connections targeting the throttle output (5 inputs plus bias).

Source	Weight	Interpretation
fatigue	+2.00	Fatigue increases throttle (fatigue-induced acceleration)
distance_to_signal	−1.91	Approaching signals reduces throttle
stress	−0.58	Stress reduces throttle (caution response)
acknowledge_aws	−2.00	AWS acknowledgement kills throttle
Hidden 353	−1.92	Arousal integrator suppresses throttle
Bias	−0.09	Near-zero baseline throttle

The throttle neuron reveals three important emergent behaviours. First, the arousal integrator (Hidden 353, weight −1.92) provides a combined fatigue-stress-braking signal that modulates throttle output. This hidden neuron compresses multiple human factors into a single arousal dimension, creating an internal representation that has no direct analogue in the sensor inputs.

Second, the signal proximity connection (distance_to_signal, weight −1.91) shows that the agent progressively reduces throttle as it approaches signals. This is not binary throttle-off behaviour; the negative weight creates a continuous reduction proportional to proximity, producing the gradual deceleration characteristic of professional driving.

Third, the AWS acknowledgement connection (weight −2.00, the maximum magnitude) functions as a throttle kill switch. When the driver acknowledges an AWS warning, throttle is immediately suppressed. This couples the warning acknowledgement to a physical speed reduction, ensuring that acknowledging a caution signal produces an actual driving response rather than a mere button press. Real-world AWS systems are designed to achieve exactly this coupling, and the evolved connectome discovered it independently.

4.4 Braking Neuron Connections

Table 6. Connections targeting the brake output (4 sensor inputs, 2 hidden inputs, plus bias).

Source	Weight	Interpretation
signal_aspect	−0.13	Weak direct signal influence (delegated to hidden neurons)
visibility	−2.00	Poor visibility strongly increases braking
route_familiarity	−2.00	Familiar routes increase braking conservatism
cognitive_load	−0.39	Higher cognitive load increases braking caution
Hidden 996	+0.56	Danger detector increases braking
Hidden 1059	−1.58	Yellow signal filter modulates braking
Bias	−0.14	Slight baseline braking tendency

The braking neuron demonstrates the most sophisticated signal processing in the evolved topology through its two-pathway architecture. Rather than connecting signal_aspect directly to braking with a single strong weight, the network routes signal information through two hidden neurons that perform distinct filtering operations.

Hidden 996 (the "danger detector") receives signal_aspect with weight −0.90 and outputs to brake with weight +0.56. Its sigmoid activation means it produces high output when signal_aspect is low (representing danger, as lower values correspond to more restrictive signals), effectively detecting red signal conditions and increasing braking force.

Hidden 1059 (the "yellow signal filter") receives signal_aspect with weight +0.62 and outputs to brake with weight −1.58. Its Gaussian activation function produces peak output at intermediate signal values (yellow and double-yellow aspects), modulating the braking response for caution signals differently from danger signals.

Together, these two pathways allow the network to discriminate between red signals (requiring maximum braking) and yellow signals (requiring moderate braking), a distinction that a single direct connection cannot achieve. The network invented multi-pathway signal discrimination to solve a problem that requires non-linear response profiles.

4.5 Emergency Brake Neuron Connections

Table 7. Connections targeting the emergency brake output (4 inputs plus bias).

Source	Weight	Interpretation
braking_distance	−1.46	Short braking distance triggers emergency brake
fatigue	−1.41	High fatigue triggers emergency brake
cognitive_load	−1.44	High cognitive load triggers emergency brake
faid_score	−0.71	Low alertness score triggers emergency brake
Bias	+0.03	Near-zero baseline (normally inactive)

The emergency brake neuron exhibits a remarkable emergent pattern: a dead man's switch. In real-world rail systems, the dead man's switch is a safety device that applies emergency braking if the driver becomes incapacitated (releases the control). The evolved connectome has discovered an analogous mechanism through a different pathway.

All four input connections carry negative weights, meaning the emergency brake activates when inputs are low or negative. The bias is near zero (+0.03), so the neuron is balanced at the activation threshold. When the driver is alert and engaged (high braking_distance indicating safe following, low fatigue, low cognitive load, high FAID score), the negative weights suppress emergency braking. But when the driver's state degrades across multiple dimensions simultaneously (high fatigue, high cognitive load, low alertness, short braking distance), the accumulated negative contributions push the neuron past its activation threshold and trigger the emergency brake.

This is functionally equivalent to a dead man's switch: the emergency brake fires not because of a specific danger signal, but because the driver's overall capacity to respond has degraded below a safe threshold. The system assumes danger when the driver can no longer be trusted to respond to it.

4.6 Hidden Neuron Summary

Table 8. All 7 hidden neurons in the evolved topology, with activation function, inputs, outputs, and functional role.

Neuron	Activation	Inputs (weight)	Output (weight)	Role
Hidden 353	Sigmoid	fatigue (−0.82), stress (−1.82), brake (+0.60)	throttle (−1.92)	Arousal integrator
Hidden 414	Sigmoid	gradient (+0.67)	acknowledge_aws (+1.67)	Gradient-AWS link
Hidden 568	Sigmoid	route_familiarity (+0.87), aws_alert (−1.05)	attention (+1.13), Hidden 702 (−0.10)	Familiarity-alert integration
Hidden 702	Gaussian	at_station (−1.80), Hidden 568 (−0.10)	acknowledge_aws (+1.67)	Station state processor
Hidden 996	Sigmoid	signal_aspect (−0.90)	brake (+0.56)	Danger detector
Hidden 1059	Gaussian	signal_aspect (+0.62)	brake (−1.58)	Yellow signal filter
Hidden 1677	Sigmoid	at_station (−1.03)	acknowledge_aws (+1.34)	Not-at-station detector

The seven hidden neurons form a processing network that operates at multiple levels of abstraction. Hidden 353 (the arousal integrator) compresses three human factors dimensions into a single signal. Hidden 568 and 702 form a two-layer pipeline for familiarity and alert processing. Hidden 996 and 1059 provide dual-pathway signal discrimination. Hidden 414 and 1677 handle AWS acknowledgement logic based on gradient and station state respectively.

Two neurons use Gaussian activation (Hidden 702 and Hidden 1059), which produces peak output at specific input ranges rather than the monotonic response of sigmoid neurons. This is particularly important for Hidden 1059's role as a yellow signal filter, where the Gaussian response peaks at intermediate signal values and drops off at both extremes (green and red). The network exploited the available activation function diversity to solve a discrimination problem that sigmoid neurons alone could not address efficiently.

5. Emergent Behaviours: What the Brain Invented

The following behaviours were not designed, rewarded, or suggested by the fitness function. They emerged solely from the selection pressure created by the attention-gated perception environment and parallel evaluation.

5.0.1 Complacency Countermeasure

The route_familiarity-to-attention connection (weight +1.56) causes the driver to pay more attention on familiar routes. This directly counteracts the complacency effect observed in human operators, where familiarity with a task leads to reduced vigilance. The evolved connectome treats familiarity as a danger signal rather than a comfort signal, increasing attentional resources when operating on well-known routes. This behaviour emerged because drivers who relaxed on familiar routes received degraded perception and crashed at locations they had previously navigated successfully, creating strong selection pressure against familiarity-induced complacency.

5.0.2 Dual-Pathway Signal Discrimination

The braking system uses two hidden neurons (Hidden 996 and Hidden 1059) to process signal information through parallel pathways, producing different braking responses for red signals versus yellow signals. Hidden 996 (sigmoid, weight −0.90 from signal_aspect) activates strongly for danger signals, while Hidden 1059 (Gaussian, weight +0.62 from signal_aspect) peaks for intermediate caution signals. This allows the network to apply maximum braking for red signals while applying proportional, moderate braking for yellow signals. A single direct connection from signal_aspect to brake cannot achieve this non-linear discrimination.

5.0.3 Dead Man's Switch

The emergency brake neuron activates when the driver's overall state degrades below a threshold across four dimensions (braking distance, fatigue, cognitive load, and FAID score). All four connections are negative, meaning the emergency brake fires when the driver is too impaired to respond to specific danger signals. This is functionally equivalent to a physical dead man's switch, which applies emergency braking when the driver releases the control handle due to incapacitation. The evolved version monitors internal state rather than physical grip, but achieves the same safety function.

5.0.4 Fatigue-Induced Acceleration

The fatigue-to-throttle connection (weight +2.00) causes fatigued drivers to accelerate. This seemingly dangerous behaviour mirrors a well-documented phenomenon in human factors research: fatigued operators tend to speed up to maintain arousal and reach their destination sooner, compensating for reduced alertness with increased pace. The behaviour persists in the evolved population because the marginal fitness gain from faster completion occasionally outweighs the crash risk, creating a stable (if risky) strategy that balances speed against safety.

5.0.5 Stress-Induced Caution

The stress-to-throttle connection (weight −0.58) causes stressed drivers to reduce speed. Unlike the fatigue response (which increases speed), stress produces the opposite effect: cautious deceleration. This distinction between fatigue and stress responses is psychologically realistic. Fatigued operators seek stimulation through speed, while stressed operators seek safety through caution. The network discovered these opposing response patterns independently, without any instruction about the psychological difference between fatigue and stress.

5.0.6 AWS Throttle Kill

The acknowledge_aws-to-throttle connection (weight −2.00) immediately suppresses throttle when the driver acknowledges an AWS warning. This ensures that acknowledging a caution signal produces an actual speed reduction rather than merely pressing a button. Real-world AWS systems are designed with exactly this intent: the acknowledgement should trigger a driving response, not just silence an alarm. The evolved connectome couples these two actions at the neural level, making throttle reduction an automatic consequence of warning acknowledgement.

5.0.7 Signal Proximity Throttle Reduction

The distance_to_signal-to-throttle connection (weight −1.91) creates a continuous, distance-proportional throttle reduction as the driver approaches any signal. This produces the smooth, gradual deceleration that characterises professional driving, rather than the abrupt throttle-off behaviour that would result from a simple threshold-based approach. The negative weight means throttle decreases as distance_to_signal decreases (as the signal gets closer), creating a natural braking curve.

5.0.8 The Arousal Integrator

Hidden 353 combines fatigue (−0.82), stress (−1.82), and braking state (+0.60) into a single arousal signal that modulates throttle (−1.92). This hidden neuron functions as an internal model of the driver's overall arousal level, compressing three distinct human factors dimensions into a one-dimensional representation. When arousal is low (high fatigue, low stress, no braking), the integrator allows throttle. When arousal is high (low fatigue, high stress, active braking), it suppresses throttle. The inclusion of braking state as an input is notable: it means the arousal integrator accounts for the driver's current actions, not just their internal state, creating a feedback loop between behaviour and arousal.

6. Discussion

6.1 Designed vs Emergent

Table 9. What was designed into the simulation versus what the evolved connectome invented.

Designed	Emergent
18 sensors exist	Which sensors matter for which actuators
5 actuators exist	When and how strongly each fires
Attention gates perception	Route familiarity increases attention
Signals have aspect values	Red and yellow require different responses
Emergency brake actuator exists	Dead man's switch pattern
Fatigue accumulates over time	Fatigued drivers speed up
Stress spikes from incidents	Stressed drivers slow down
AWS horn + acknowledgement system	AWS acknowledgement kills throttle
Signals exist along the track	Approaching signals reduces throttle
Fatigue, stress, braking are separate sensors	Arousal integrator combining all three

The designed/emergent distinction is fundamental to understanding the Quale approach. The simulation provides the raw materials (sensors, actuators, environmental dynamics) but does not specify how they should be used. Every behavioural strategy listed in the "Emergent" column was discovered by the NEAT process through selection pressure alone. The simulation designer did not encode that familiarity should increase attention, that red and yellow signals need different responses, or that fatigue should cause acceleration. These are discoveries, not implementations.

6.2 Depth Enables Discrimination

The most significant topological difference between Rail-003 and Rail-003b is the evolution of deep processing pathways. Rail-003's single functional hidden neuron performed a straightforward integration (speed, distance, signal state combined into a station approach assessment). Rail-003b's 5+ functional hidden neurons form multi-layer processing pipelines that enable qualitatively more sophisticated computation.

The clearest example is dual-pathway signal discrimination. A single direct connection from signal_aspect to brake can only produce a monotonic response (more restrictive signals produce more braking, linearly). The two-pathway architecture (Hidden 996 and Hidden 1059) allows non-linear, aspect-specific responses: maximum braking for red, moderate braking for yellow, minimal braking for green. This discrimination is essential for realistic driving, where the response to a yellow signal (prepare to stop) is fundamentally different from the response to a red signal (stop immediately).

Parallel evaluation drove this depth because genomes were tested across diverse signal configurations within each generation. A shallow network that brakes identically for yellow and red signals will underperform on routes where yellow signals are common (excessive braking wastes time) and on routes where red signals follow closely after yellows (insufficient braking causes violations). Only a deep network with non-linear discrimination can perform well across both scenarios.

6.3 Complacency Countermeasure

The route_familiarity-to-attention connection (weight +1.56) is perhaps the most psychologically significant finding in the Rail series. In human factors research, complacency is one of the most persistent safety challenges: operators who know a task well tend to reduce their vigilance, leading to errors that novice operators would not make. Training programmes spend considerable effort teaching operators to resist this tendency.

The evolved connectome solved this problem without any knowledge of human psychology. It simply discovered that drivers who maintained high attention on familiar routes survived more consistently than those who relaxed. The positive weight emerged because the attention gate penalises inattention regardless of route familiarity, and familiar routes provide a false sense of security that leads to degraded perception and eventual crashes.

This finding suggests that complacency is not merely a human psychological weakness but a fundamental failure mode of any cognitive system operating under attention-gated perception. Any agent that reduces vigilance based on familiarity will eventually fail, because familiarity does not eliminate environmental hazards; it only reduces the perception of those hazards.

6.4 Dead Man's Switch

The emergency brake's dead man's switch pattern demonstrates that safety-critical behaviours can emerge without explicit safety engineering. The physical dead man's switch was invented by railway engineers as a mechanical failsafe: if the driver becomes incapacitated, the control handle is released, and the brakes apply automatically. The evolved connectome achieves the same functional outcome through a completely different mechanism: monitoring the driver's internal state across multiple dimensions and triggering emergency braking when overall capacity degrades below a threshold.

The evolved version is, in some respects, more sophisticated than the mechanical original. A physical dead man's switch only detects complete incapacitation (hand off the control). The evolved version detects gradual degradation across fatigue, cognitive load, and alertness dimensions, potentially triggering emergency braking before the driver becomes fully incapacitated. This is analogous to modern driver monitoring systems that track eye movements and reaction times to detect drowsiness before the driver falls asleep.

6.5 Fatigue-Speed Paradox

The fatigue-to-throttle connection (weight +2.00, increasing speed with fatigue) appears dangerous, yet it persists in the evolved population because it represents a locally optimal strategy. Fatigued drivers who maintain or increase speed complete route segments faster, accumulating less total fatigue exposure and reaching stations (where fatigue resets) sooner. The alternative (slowing down when fatigued) extends the time spent in a degraded state, potentially leading to more errors overall.

This is consistent with real-world observations. Fatigued drivers report speeding up to "stay awake" or "get there faster," and studies show that fatigue-related speeding is one of the most common fatigue-induced behaviours. The evolved connectome discovered this strategy independently, suggesting that it represents a genuine (if risky) optimisation under fatigue conditions rather than merely a human cognitive bias.

The counterbalance to this dangerous strategy is the dead man's switch: if fatigue accumulates too far, the emergency brake fires regardless of speed. This creates a bounded risk profile where fatigue-induced acceleration operates within a safety envelope defined by the emergency brake threshold.

6.6 Parallel Evaluation Impact

The comparison between Rail-003 (sequential) and Rail-003b (parallel) reveals that evaluation mode has a profound effect on evolved topology and behaviour. Parallel evaluation produced deeper networks (7 hidden neurons versus 2 in Rail-003, of which only 1 was functional), more connections (36 versus 22), lower idle rates (48% versus 71%), and higher final fitness (99.69 versus 99.03). These differences are not marginal; they represent a qualitative shift in the type of solution that evolution discovers.

The mechanism is straightforward: parallel evaluation tests each genome across multiple scenarios within a single generation, penalising specialisation and rewarding generalisation. A shallow network that performs well on one route configuration but poorly on another will receive mediocre average fitness and be outcompeted by deeper networks that handle both configurations. This drives the evolution of deeper topologies with context-sensitive responses, because depth is the topological mechanism that enables context sensitivity.

The wall-clock time reduction (from approximately 35 minutes to 70 seconds) demonstrates that parallel evaluation is not merely a methodological improvement but a practical enabler of deeper evolution. Experiments that would take hours under sequential evaluation can be completed in minutes, enabling rapid iteration and exploration of the parameter space.

7. Conclusion

1. Parallel evaluation drives deeper topology evolution. The shift from sequential to parallel evaluation produced a genome with 7 hidden neurons and 36 connections, compared to Rail-003's 2 hidden neurons (1 functional) and 22 connections, with higher final fitness (99.69 versus 99.03).
2. Route familiarity emerges as a complacency countermeasure. The evolved connectome increases attention on familiar routes (weight +1.56), independently discovering a principle that human factors training programmes teach explicitly.
3. Dual-pathway signal discrimination evolves from depth. Two hidden neurons with different activation functions (sigmoid and Gaussian) process signal_aspect through parallel pathways, enabling non-linear discrimination between red and yellow signal responses.
4. A dead man's switch emerges from multi-dimensional state monitoring. The emergency brake activates when the driver's overall capacity degrades across fatigue, cognitive load, braking distance, and alertness dimensions, mirroring the function of mechanical dead man's switches in real-world rail systems.
5. Fatigue-induced acceleration is a stable evolutionary strategy. The fatigue-to-throttle connection (weight +2.00) produces a risky but locally optimal behaviour, bounded by the emergency brake safety envelope, that mirrors documented human fatigue responses.
6. The arousal integrator compresses multiple human factors into a single dimension. Hidden 353 combines fatigue, stress, and braking state into a unified arousal signal, demonstrating that NEAT can evolve internal representations that have no direct analogue in the sensor space.
7. Parallel evaluation is both faster and better. The 30x wall-clock speedup (70 seconds versus 35 minutes) combined with higher fitness and deeper topology suggests that parallel evaluation should be the default for future Rail experiments.

8. Cross-Experiment Summary

Table 10. Summary of all four Rail experiments across 12 metrics.

Metric	Rail-001	Rail-002	Rail-003	Rail-003b
Throttle strategy	Binary on/off	Speed-limited	Hidden-neuron modulated	10-input continuous control
Attention strategy	None	None	3-input context-sensitive	4-input with familiarity countermeasure
Evaluation mode	Sequential	Sequential	Sequential	Parallel
Best fitness	156	389	551 (99.03 normalised)	99.69 (normalised)
Hidden neurons	0	0	1 functional	5+ functional (7 total)
Connections	8	13	22	36
Minimum idle rate	72%	71%	71%	48%
Signal processing	None	Direct connection	Single pathway	Dual-pathway discrimination
Complacency countermeasure	No	No	No	Yes (route familiarity +1.56)
Dead man's switch	No	No	No	Yes (4-input threshold)
Human factors parallels	Basic survival	Speed compliance	Yerkes-Dodson law	Complacency, fatigue acceleration, arousal integration
Wall-clock time	~10 min	~20 min	~35 min	70 sec