⚠ AI-Driven Alpha Prototype — Experimental. May be misleading. For exploration only. ← Back to demos
Infancy Phase — BDH·DLiNoSS·HFF PID Brain
?INFANCY PHASE SIMULATION
A dual-cage environment for evolving water-skidder
creatures with merged BDH+DLiNoSS+HFF PID brains.
Each cage contains a winged stickman body controlled
by a tiny evolutionary neural controller that learns
reflexive behaviors through reward-modulated Hebbian
learning, oscillator banks, and forward-forward
prediction.
GOOD SIGNS: Smooth swimming motions, ripple patterns
in the tank, reward trending upward, low jitter %.
BAD SIGNS: Frozen creatures, >50% jitter, reward
stuck negative, creatures pinned to walls.
Drag particles to interact. Use sliders to tune.
Infant Goal Weights
?INFANT GOAL WEIGHTS
Control the reward function that drives learning.
Each weight scales a reward/penalty component in
the fitness evaluation. Higher = more influence.
These are the most important tuning parameters —
they determine WHAT the creature learns to do.
GOOD SIGN: Balanced weights producing varied
smooth behaviors. Net reward trending upward,
creature exploring its space.
BAD SIGN: One weight dominating others causing
stuck/repetitive behavior. Net reward flat at
zero or negative for many generations.
TIP: Start with defaults, then increase 'Wallow'
for movement exploration or 'Stretch' for wing
extension practice.
Adjust gradually — large jumps reset learning.
Don't Hurt
?DON'T HURT — Pain Penalty Weight
Scales how strongly wall-collision pain penalizes
the creature's reward. High values make the creature
very pain-averse, low values let it explore walls.
GOOD SIGN: Creature avoids walls but still moves
freely. Pain level near zero most of the time.
BAD SIGN: Set too high and creature freezes in
center (over-cautious). Set too low and creature
repeatedly smashes into walls.
The pain signal is a 3-timescale system:
instant (contact), short-term (fading), long-term
(cumulative). All feed into the penalty.
Default: 1.00. Range: 0–2.
High pain-aversion (>1.5) can suppress exploration.
Zero means the creature ignores all pain.1.00
Wall Avoid
?WALL AVOID — Proximity Gradient Penalty
Penalizes the creature for being close to cage
walls, even without touching them. Creates a
gradient that pushes the creature inward.
Only activates when wall proximity > 0.6 (inner
40% of the cage is a safe zone).
GOOD SIGN: Creature stays in the central region
with occasional wall approaches.
BAD SIGN: Too high and creature is trapped in a
tiny center zone. Too low and it hugs walls.
Works as a soft boundary — complementary to
Don't Hurt which is a hard collision penalty.
Default: 0.25. Range: 0–2.
Values above 1.0 create strong repulsion.
Combine with Center weight for balanced positioning.0.25
Center
?CENTER SEEK — Centering Reward
Rewards the creature for positioning its center
of mass near the horizontal center of the cage.
Reward = 1 − (distance / half_cage_width).
GOOD SIGN: Creature tends toward center but
still explores outward occasionally.
BAD SIGN: Too high = creature stays perfectly
still at center (no exploration). Too low =
creature drifts to edges.
Works with Wall Avoid as a push-pull pair:
Wall Avoid repels from edges, Center attracts
to center. Balance them carefully.
Default: 0.15. Range: 0–2.
Typical useful range: 0.05–0.50.
Set to 0 for free-roaming behavior.0.15
Hiccup
?HICCUP — Periodic Twitch Reward
Rewards the creature for making periodic body
twitches at roughly 1.2 Hz (matching an internal
sinusoidal reference). Encourages rhythmic muscle
contractions like fetal hiccups.
GOOD SIGN: Visible periodic body contractions,
creature develops a rhythm. The oscillator bars
show synchronized patterns.
BAD SIGN: Too high = creature only twitches and
never develops smooth movement. Too low = missed
opportunity for rhythmic motor pattern learning.
This is a developmental milestone — real fetuses
hiccup to train respiratory muscles.
Default: 0.10. Range: 0–2.
Keep low (0.05–0.20) for subtle rhythmic training.
Increase to encourage musical body oscillation.0.10
Stretch
?STRETCH — Limb Extension Reward
Rewards the creature for extending its limbs
(sticks) toward their rest length. Measures the
ratio of current length to rest length across
all body sticks.
GOOD SIGN: Wings and spine fully extended, body
adopts an open posture. Stick stretch ratios
near 1.0 in the sensor readout.
BAD SIGN: Too high = creature strains to over-
extend, possible instability. Too low = creature
curls up into a ball.
Stretch + Wallow together encourage graceful
swimming motions — stretch opens the body,
wallow rewards forward gliding.
Default: 0.15. Range: 0–2.
Works well in the 0.10–0.40 range.
Increase for more wing-spread behaviors.0.15
Wallow
?WALLOW — Gentle Movement Reward
Rewards sustained, gentle movement (not too fast,
not still). Uses a Gaussian bell curve centered
at ~0.5 px/frame — the sweet spot of graceful
wallowing motion.
GOOD SIGN: Creature drifts smoothly through the
medium with consistent slow speed. Beautiful
ripple patterns in the tank.
BAD SIGN: Too high = creature optimizes for
constant speed and ignores other goals. Too low =
creature is either frozen or frantic.
Key reward for developing swimming behavior in
the aquatic medium. Combines naturally with
Stretch for elegant water-skidder locomotion.
Default: 0.20. Range: 0–2.
The most important movement reward. Start here
when tuning for water-skidder behavior.0.20
Withdraw
?WITHDRAW — Pain Avoidance Reward
Rewards the creature for moving AWAY from the
last wall it collided with. Teaches the
association: pain → move in opposite direction.
Only activates after a wall hit occurs, and
measures whether the COM velocity points away
from the hit wall.
GOOD SIGN: After a wall bump, creature smoothly
reverses direction. Quick recovery from pain.
BAD SIGN: Too high = creature overreacts to
wall contact with violent jerks. Too low =
creature stays near walls after collision.
A fundamental reflexive behavior — fetal
withdrawal reflex. Pairs with Don't Hurt for
complete pain-response learning.
Default: 0.30. Range: 0–2.
Keep moderate (0.20–0.50) for natural reflex.
High values create twitchy wall-avoidance.0.30
Presets
?PRESET SLOTS — Save & Recall Settings
8 slots to save and recall all your slider settings
(physics parameters, goal weights, speed).
CLICK empty slot: Saves current settings there.
CLICK occupied slot (green): Loads those settings.
RIGHT-CLICK occupied slot: Clears that slot.
Presets persist in your browser's localStorage
across page reloads.
GOOD FOR: A/B comparison of configurations,
saving a good baseline before experimenting,
quickly switching between different training
regimes (e.g., "explore" vs "settle" presets).
The [Defaults] button resets all sliders to the
original factory values without affecting presets.
TIP: Save a preset before making major changes
so you can always revert.
Physics Tuning
?PHYSICS TUNING
Adjust the Verlet physics engine parameters that
govern how the creature's body behaves in the
aquatic medium. Changes apply immediately.
These parameters are your main tool for fixing
physics instability (jitter/tremor/twitching).
GOOD SIGN: Smooth particle motion, sticks at
rest length, jitter indicator at 0%.
BAD SIGN: Particles vibrating, sticks stretching
wildly, jitter >20%, body exploding.
PRIMARY JITTER FIX: Increase Mass (adds inertia)
and increase Damping (absorbs oscillation).
SECONDARY: Decrease MaxSpd and Bounce.
All changes are included when saving presets.
TIP: If the body is jittering, try Mass=5+
and Damping=0.97+ first.
Mass
?MASS — Particle Inertia
Sets the mass of all body particles. Higher mass
= more inertia = smoother but slower motion.
Mass directly affects Newton's F=ma — higher
mass means the same torque produces less
acceleration, reducing jitter and overshoot.
GOOD SIGN: Body moves smoothly with visible
momentum. Natural-looking inertial gliding.
BAD SIGN: Too high = sluggish, unresponsive body
that the brain struggles to move. Too low =
twitchy, jittery motion from tiny forces.
This is the #1 parameter for fixing jitter.
Real water insects have surprisingly high mass
relative to their muscle force.
Default: 3.0. Range: 0.5–10.
Jitter fix: try 5.0–8.0.
Light/fast: 1.0–2.0 (may jitter).3.0
Damping
?DAMPING — Velocity Retention
Fraction of velocity retained each frame.
0.96 means 4% velocity lost per step.
Lower = more drag = faster stop.
Acts like viscous drag in the amniotic fluid.
Critical for stability — too little damping
and oscillations grow, too much and the
creature cannot move.
GOOD SIGN: Body glides to a smooth stop.
Oscillations die out within a few frames.
BAD SIGN: Too high (>0.99) = oscillations ring
forever. Too low (<0.90) = body barely moves,
feels stuck in molasses.
Works with Mass: high mass + high damping =
very stable but slow. Low mass + low damping =
responsive but potentially unstable.
Default: 0.96. Range: 0.80–0.999.
Jitter fix: try 0.95–0.97.
Aquatic feel: 0.94–0.96.0.96
MaxSpd
?MAX SPEED — Velocity Clamp
Hard limit on particle velocity (px/frame).
Prevents particles from moving too fast and
causing physics explosion or tunneling through
cage walls.
GOOD SIGN: Speed never hits the limit. Body
moves freely within natural speed range.
BAD SIGN: Particles constantly hitting speed
cap = forces are too strong for current mass.
Increase mass or decrease torque instead.
This is a safety valve, not a tuning parameter.
If you need to lower this to prevent explosion,
the real fix is usually more damping or mass.
Default: 40. Range: 5–200.
Conservative: 20–30 (safe, limits exploration).
Normal: 40–80.
High risk: 100+ (may cause instability).40
Bounce
?BOUNCE — Wall Collision Elasticity
How much velocity is reflected when a particle
hits a cage wall. 0 = perfectly inelastic
(stops dead), 0.5 = half velocity reflected.
GOOD SIGN: Soft wall contacts, creature slides
along walls without bouncing away violently.
BAD SIGN: High bounce causes the creature to
oscillate between walls (ping-pong effect).
Also worsens jitter near walls.
Low bounce (0–0.05) is better for aquatic feel.
Real water surfaces have very low elasticity.
Default: 0.05. Range: 0–0.5.
Recommended: 0–0.08 for water-skidder.
Higher values create more energetic collisions.0.05
W.Drag
?WATER DRAG — Medium Resistance
Additional directional drag simulating water
resistance on particle surfaces. Adds realistic
aquatic friction beyond the global damping.
GOOD SIGN: Body movements feel like swimming
in a viscous medium. Natural deceleration
curves.
BAD SIGN: Too high = body freezes, cannot
overcome water resistance. Too low = body
moves like it's in vacuum (unrealistic).
Combines with Damping: damping is a global
velocity multiplier, W.Drag is extra drag
proportional to speed in each direction.
Default: 0.06. Range: 0–0.3.
Light water: 0.02–0.05.
Heavy medium (amniotic): 0.06–0.12.
Thick gel: 0.15+ (very slow).0.06
Tether
?TETHER — Center Pull Strength
A weak spring force pulling all particles toward
the cage center. Prevents the creature from
permanently lodging itself in a corner.
GOOD SIGN: Creature naturally tends toward
center but can still explore edges freely.
BAD SIGN: Too high = creature is rubber-banded
to center, cannot explore. Too low = creature
drifts to edges and stays there.
This is a very subtle force — keep it low.
It should be barely perceptible, just enough
to break symmetry when the creature is stuck.
Default: 0.005. Range: 0–0.05.
Subtle: 0.001–0.005.
Noticeable: 0.01–0.02.
Strong (not recommended): 0.03+.0.005
Paddle
?PADDLE — Wing Thrust Force
Maximum thrust force the wing paddles can exert
when activated by the brain. Higher = more
powerful wing strokes, more ripple disturbance.
GOOD SIGN: Wings produce visible ripples in
the tank, creature can propel itself through
the medium with paddle strokes.
BAD SIGN: Too high = violent wing strokes
cause body instability and jitter. Too low =
wings flap but produce no useful thrust.
This is the main locomotion force for the
water-skidder concept. Balance with Mass:
Paddle/Mass ratio determines acceleration.
Default: 180. Range: 10–500.
Gentle: 50–100 (subtle movement).
Normal: 150–250 (visible propulsion).
Powerful: 300+ (may cause instability).180
Brain & Physics Analytics
?REAL-TIME ANALYTICS VIEWER
Zoomable, pannable view of ALL internal brain and
physics signals with live FFT spectrograms and
VU-bar level meters.
MOUSE CONTROLS:
• Scroll wheel: zoom in/out (toward cursor)
• Left-drag: pan the view
• Right-click: choose a preset view
BUTTONS:
• Home: reset zoom/pan to default
• Fit: auto-fit all content to screen
DOWNSAMPLING:
When running at high speed, the downsampler
prevents aliasing artifacts in the FFT display.
"Skip" is fast (just decimates), "AA" applies
a box filter for accurate spectral analysis.
15 preset views cover every aspect of the brain
constellation: DLiNoSS oscillators, BDH Hebbian,
HFF forward-forward, PID gains, reward breakdown,
wing dynamics, and more.
?HOME resets pan & zoom to the default origin.
All signal strips return to their natural position.?FIT auto-scales to show all visible content
within the analytics canvas.?PRESETS: Switch between signal groups
(FFT strips, VU bars, neural graph) and
3D trail views (waterfall, stack, orbit).
Waterfall: signals recede into vanishing point
Stack: layered 3D waveforms with filled shadows
Orbit: phase-space XY trajectories
Also accessible via right-click context menu.?DOWNSAMPLING factor (1-64).
At high sim speed, reduces the number of
samples stored in the ring buffer.
Factor N means keep 1 in every N samples.
Prevents FFT aliasing at >100x speed.?PHASE offset for skip-mode downsampling.
When DS=4 and Phase=2, we keep sample
2, 6, 10, 14... instead of 0, 4, 8, 12...
Different phases reveal different aliasing
artifacts — sweep through to compare.
(Only active in Skip mode, not AA mode.)?DOWNSAMPLE MODE:
Skip: fast decimation, keeps every Nth sample.
May introduce aliasing (educational!).
Phase slider available to shift the sample pick.
AA: anti-aliased box filter averaging.
More CPU but spectrally clean.
No phase slider (all samples contribute).
SIM HEALTH:
3D Neural Graph
?INTERACTIVE 3D NEURAL VISUALISATION
All ~118 neurons shown as spheres in force-directed
3D layout. Layers: Input → DLiNoSS → BDH → HFF →
PID-mix → Output.
Animated flow dots show spike events (orange) and
sampling streams (blue) along connections.
MOUSE CONTROLS:
• Left-drag: orbit / rotate
• Scroll wheel: zoom in/out
• Right-drag: pan
Node colour = PID channel assignment.
Brightness pulses on spike detection.
DRAG a node to pluck-learn: perturbs brain weights!
3D Usability Layout Monitor
?INTERACTIVE USABILITY ERGONOMICS MAP
Force-directed 3D graph of ALL interactive UI elements.
HOW IT WORKS:
• Every click, slider drag, or select change is tracked
• Mouse travel distance between interactions is recorded
• Items used one after another attract each other
• High mouse travel distance pushes items apart
NODE BRIGHTNESS: more used = brighter glow
NODE SIZE: proportional to total usage count
EDGES: connect items used in sequence — thicker
edges = more frequent transitions
This is a gravitational UI optimiser — frequently
co-used controls naturally cluster together,
revealing ergonomic improvement opportunities.
DRAG nodes to manually explore layouts.
The graph self-organises from your actual usage patterns.
GOAL: items that are clicked together should be
neighbours — minimising cognitive navigation load.
AI Assistant
?EMBEDDED MCP-AWARE AI ASSISTANT
Continuous-learning assistant that tracks everything:
HOW IT WORKS:
• Every interaction, parameter change, and conversation
is logged as JSONL checkpoint events with Unix timestamps
• Knowledge accumulates from your usage patterns
• Conversations are correlated with tracked interactions
• Model fingerprints ensure parameter consistency
CHECKPOINTING (Ⓒⓚ2026afrtNL):
• ⬇ Export: Download a .jsonl checkpoint file
• ⬆ Import: Load a previously exported checkpoint
• Server-side merge: Combine multiple checkpoints
using npx tsx merge-checkpoints.ts
JSONL FORMAT (MCP Protocol):
Each line is a jsonrpc 2.0 message with:
timestamp, copyright, model fingerprint, type, payload
The assistant learns WHY and WHAT changed by correlating
conversations with interactions over time windows.
Spectral Learning EQ
?SPECTRAL LEARNING EQUALISER
Three-band control over the learning rate multiplier.
DC (bias): Constant multiplier applied to all
BDH Hebbian and HFF forward-forward learning.
0 = no learning, 1 = normal, 2 = double speed.
1st Order (gradient): Scales learning by the
first derivative (trend) of recent reward.
Positive = boost learning when reward is rising.
Negative = boost when reward is falling.
2nd Order (curvature): Scales learning by the
second derivative (curvature) of recent reward.
Positive = boost at reward acceleration.
Negative = boost at reward deceleration.
Combined: mul = DC + 1st·gradient·10 + 2nd·curvature·50
Clamped to [0, 4] for safety.