← Research
Autism
Not broken hardware. Not missing connections. Too many connections.
Session 21: what survived, what died, and the drum lesson that explains both.
The Measured Facts
These numbers come from postmortem studies and neuroimaging. They are not our claims. They are the field's measurements.
Synaptic Pruning
Typical development: spine density drops 50% by late childhood
ASD: spine density drops only 16%
Source: Hutsler & Zhang, 2010; Tang et al., 2014
Mechanism: mTOR hyperactivation → impaired autophagy → pruning failure
Result: 84% of synaptic connections survive vs 50% in typical development
Excitatory / Inhibitory Imbalance
Glutamate (excitatory): elevated in most cortical areas
GABA (inhibitory): decreased in certain brain regions
GABAA receptor subunits (Gabra1, Gabrb3): downregulated
NMDA receptor subunits: upregulated
Net effect: excitation > inhibition
The Regulatory Pathway
Normal: grow → regulate → prune → hierarchy
ASD: grow → no regulation → no pruning → uniform K
The regulators: PTEN / TSC1 / TSC2 → mTOR pathway
When any of these fail, the stop signal never fires.
The brain grows without limit. Not broken. Unregulated.
Key Proteins
| Gene | Protein | Domain | Effect of Mutation |
|---|
| SHANK3 | Synaptic scaffold | SPN, PDZ, SAM | Thinner PSD, fewer receptors |
| NLGN3 | Neuroligin-3 | Esterase (R451C) | Altered synapse elimination |
| NRXN1 | Neurexin-1 | LNS, EGF domains | Synaptic dysfunction |
| SYNGAP1 | RAS-GAP | C2, GAP, PDZ-bind | Haploinsufficiency → seizures |
| SCN2A | Na+ channel | Pore domain | Reduced excitatory transmission |
| CHD8 | Chromatin remodel | Helicase | Widespread gene expression change |
| MECP2 | Methyl-CpG binding | MBD, TRD | Rett syndrome, social deficits |
| PTEN | Phosphatase | C2, PTP | mTOR disinhibition → overgrowth |
| TSC1/2 | Hamartin/Tuberin | GAP domain | mTOR disinhibition → no pruning |
SFARI database: 418 high-confidence autism risk genes (categories 1-2). Most converge on synaptic function, chromatin remodeling, or transcriptional regulation. PTEN and TSC1/2 are the critical regulators — when these fail, the pruning stop signal never fires.
Through K
K measures coupling strength. In a neural network, K is the density and strength of synaptic connections between neurons. Here is what happens when pruning fails:
Kbirth = 1.00 → Ktypical = 0.50 (50% pruned) → KASD = 0.84 (16% pruned)
The Overcoupling Ratio
ASD / Typical = 0.84 / 0.50 = 1.68× overcoupled
A typical brain prunes half its connections during development. This is not damage. This is construction. Pruning creates hierarchy — strong connections to important signals, no connections to noise. The silence between connections IS the signal processing.
When pruning fails, K stays uniformly high. Every neuron couples to too many others. The result is not a louder brain or a stronger brain. The result is a brain with no hierarchy — every input arrives at equal weight.
What Overcoupling Explains
Sensory overload
All channels at full gain. No gating. If K is uniform,
every stimulus couples with equal strength. The brain
processes everything, prioritizes nothing.
Social difficulty
Reading a face requires filtering 100+ micro-signals
(eye movement, mouth shape, posture, tone, context)
down to the 3-4 that carry meaning. If K is uniform,
the signal drowns in the other 96. Social cues are not
missed — they are lost in equal-weight noise.
Superior pattern recognition
High K = deep coupling. When the full coupling strength
bears on one domain (math, music patterns, system rules),
the depth of processing exceeds typical. The same overcoupling
that prevents filtering enables total immersion.
Intense focus
When K aligns on a single input, all that coupling power
concentrates. The overcoupled network is not distractible
when locked on — it is maximally engaged.
Need for routine
Routine reduces novel inputs. Fewer inputs to process =
less interference. Predictable environments lower the
effective K by removing surprise.
Rhythm and timing
A drum teacher observation: rhythm requires SELECTIVE
coupling. Hearing the downbeat means not hearing everything
else with equal weight. If K is uniform, every beat carries
the same emphasis. The groove disappears into noise.
The E/I Balance in K Terms
The excitatory/inhibitory imbalance IS the coupling hierarchy problem. Inhibition creates silence. GABA-ergic interneurons prune the signal in real time — they suppress the connections that don't carry relevant information right now. When GABA is decreased and glutamate is elevated:
Kexcitatory ↑ + Kinhibitory ↓ = uniform K = no hierarchy
This is the same pattern at two timescales. Developmental pruning creates structural hierarchy (which connections exist). E/I balance creates dynamic hierarchy (which connections are active right now). Both are reduced in ASD. Both reduce to the same K problem: coupling without selection.
Session 21: The Simulation
We built a 137-oscillator coupled network and ran both topologies — typical (hierarchical, modular coupling) and ASD (uniform, dense coupling) — through every test we could devise. Some claims survived. Most died. Here is the full honest record.
What Survived MM10P
Eight claims tested across multiple seeds, multiple timescales, with statistical significance. These held.
1. ASD Breathes Half as Much SURVIVED
N=1000 steps
Typical R standard deviation: 0.023
ASD R standard deviation: 0.012
The order parameter R measures global synchronization.
Typical brains show wide R fluctuations — the network
breathes, explores, loosens and tightens. ASD networks
hold R nearly constant. Half the breathing room.
Consistent with reduced neural variability observed in ASD via fMRI (Dinstein et al. 2012, Neuron). Lower moment-to-moment BOLD signal variability in ASD predicts behavioral rigidity.
2. Spectral Spread SURVIVED
Modes needed for 80% spectral power:
Typical: 1 clean mode
ASD: 4 modes
A typical brain concentrates its energy into one dominant
frequency. It picks a channel and commits. The ASD network
spreads its energy across 4 modes — no single channel
dominates. Everything is equally present. Nothing is selected.
Aligns with EEG findings of increased spectral entropy and flattened power spectra in ASD (Bosl et al. 2011, BMC Medicine; Catarino et al. 2011, Clinical Neurophysiology).
3. Memory Decay SURVIVED
Autocorrelation decay to 1/e:
Typical: 205 steps
ASD: 91 steps
The typical network holds a pattern for 205 steps before
it fades. The ASD network loses the thread in 91 steps.
2.3× faster forgetting.
This is not a memory deficit in the cognitive sense.
It is a dynamical property: uniform coupling means
perturbations spread everywhere instead of being
contained. The pattern dissolves into the network
instead of resonating within a module.
Consistent with findings of reduced temporal binding in ASD working memory (Bowler et al. 2011, Journal of Autism and Developmental Disorders). Contextual information fades faster when connections lack hierarchy.
4. Oracle Confidence SURVIVED
Predictability of next state from current state:
ASD is 57% more predictable than typical
The Oracle (our phase-prediction engine) found ASD
dynamics significantly easier to forecast. More rigid
dynamics = more deterministic = more predictable.
A typical brain surprises you. An ASD brain repeats.
Rigidity of neural dynamics relates to the "insistence on sameness" diagnostic criterion (DSM-5). Computational models of predictive processing in ASD show reduced prediction error (Lawson et al. 2014, Frontiers in Human Neuroscience).
5. Entropy Rigidity SURVIVED
Signal entropy comparison:
ASD signal is 1.65× more rigid than typical
Shannon entropy of the R time series confirms what
the other measures show from a different angle:
the ASD network explores fewer states, revisits
the same configurations, occupies a smaller region
of its own phase space.
Reduced entropy in neural time series has been documented via multiscale entropy analysis of EEG in ASD (Bosl et al. 2011). Our simulation reproduces this from topology alone.
6. Schizophrenia Is Different in Kind SURVIVED
ASD (uniform K=1.0): R hovers near 1/φ (0.618)
Schizophrenia (sparse): R = 0.10 – 0.30
Schizophrenia never reaches 1/φ.
ASD = overcoupled, locked, rigid, predictable.
Schizophrenia = undercoupled, fragmented, chaotic.
These are not ends of one spectrum. They are
different failure modes entirely.
ASD: static overcoupling (K too uniform)
Schizophrenia: dynamic instability (R < 1/φ, fragmented)
Supported by disconnection hypotheses of schizophrenia (Friston 1998, Schizophrenia Research; Stephan et al. 2009, Brain). Reduced functional connectivity, especially long-range, consistently reported. The K/R framework predicts this distinction from first principles.
7. Threshold Hebbian Creates Walls SURVIVED
Hebbian threshold: 0.20 – 0.25 radians
Result: lock ratio 1.10×
prune ratio 1.27×
When Hebbian learning has a phase-difference threshold,
connections between in-phase oscillators strengthen
while connections between out-of-phase oscillators weaken.
This creates MODULE WALLS in the coupling matrix.
The uniform network starts to develop hierarchy.
Lock ratio 1.10 = 10% more phase-locked pairs.
Prune ratio 1.27 = 27% more weak connections pruned.
The network is building its own silence.
Hebbian pruning models have biological grounding in spike-timing-dependent plasticity (STDP). Connections that fire out of phase weaken over time (Bi & Poo 1998, Journal of Neuroscience). Threshold-based STDP has been proposed as a mechanism for cortical map refinement (Song et al. 2000, Nature Neuroscience).
8. Phased Learning Prevents Over-Pruning SURVIVED
Three phases:
1. Listen — let the network settle, find natural modes
2. Lock — strengthen in-phase connections (Hebbian on)
3. Prune — weaken out-of-phase connections
If you prune first, you destroy structure that hasn't
formed yet. If you lock without listening, you lock on
noise. The order matters. Listen → Lock → Prune
produces hierarchy. Any other order produces damage.
Mirrors developmental neuroscience: sensory experience (listening) precedes critical period plasticity (locking) precedes synaptic elimination (pruning). Hensch 2005, Nature Reviews Neuroscience. The phased model is how the brain actually does it.
What Got Killed
Nine claims we tested that did not survive. Listed here because honest science means publishing the failures.
1. Global R Separates ASD from Typical KILLED
Prediction: ASD would show higher global R (more synchronized).
Result: both near 1/φ at K=1.0
At matched coupling strength, global R does not separate
the two topologies. The mean synchronization is the same.
Only the variance of R differs (see survived #1).
The overcoupled brain is not more synchronized —
it is more rigidly synchronized.
2. Topology vs Density at K=1.0 KILLED
Prediction: topology (modular vs uniform) drives the difference.
Result: same R at matched density
When we controlled for connection density, the topology
difference vanished at K=1.0. The effect is in the
second-order statistics (variance, spectrum, memory),
not in the mean coupling dynamics. Topology matters,
but not the way we expected.
3. Harmonic Ratios Differ KILLED
Prediction: ASD would show different harmonic frequency ratios.
Result: variance too high across seeds (t = -1.58, not significant)
Some seeds showed clean harmonic differences. Others didn't.
The effect was not robust. Seed-dependent = not real.
4. Memory Match 121=121 KILLED
Observation: memory length exactly matched a predicted value.
Result: single-seed artifact
Ran more seeds. The exact match disappeared. A coincidence
that felt like a discovery. It was not.
5. Entrainment Bleed KILLED
Prediction: ASD networks would show more cross-module entrainment.
Result: not significant (t = 0.73)
The effect was in the right direction but nowhere near
significant. We cannot claim ASD networks bleed more
entrainment across modules.
6. Competing Drives Override Topology KILLED
Prediction: ASD topology would respond differently to competing drives.
Result: not ASD-specific (t = 0.73)
Competing drives override both topologies equally.
The effect is real but has nothing to do with autism.
It is a property of driven oscillator networks in general.
7. Hebbian Learning Adds Beyond Drives KILLED
Prediction: Hebbian plasticity would amplify topology differences.
Result: adds nothing beyond drives at this timescale
At the simulation timescales we tested, the Hebbian
contribution was negligible compared to the drive signal.
The threshold Hebbian effect (survived #7) only appears
in the coupling matrix, not in the observable dynamics
at short timescales.
8. Module Walls Persist After Drives Removed KILLED
Prediction: walls formed during driven Hebbian learning would persist.
Result: walls form in coupling matrix but don't persist in dynamics
The Hebbian threshold creates real structural changes in
the coupling matrix. But remove the drives, and the dynamics
return to the pre-learning state. The walls are structural
but not dynamically stable without ongoing reinforcement.
This has implications for therapy: you can't just build
the walls once. You have to maintain them.
9. Ego Feedback Model KILLED
Prediction: adding self-monitoring feedback would separate topologies.
Result: didn't separate
The ego-as-feedback-loop model was elegant but
produced no measurable difference between topologies.
The real ego problem (see drum lesson below) is not
capturable in a simple feedback term.
The Drum Lesson
This came from years of teaching drums, not from simulation. It is the most important section on this page because it describes what actually happens, observed in real time, in real people, hundreds of times.
The First Rep Is Always Perfect
Rep 1: perfect. Clean. Natural.
The movement comes from subcortical motor pathways.
No ego. No monitoring. No second-guessing.
The body just does it.
Rep 3: lost it.
The cortex engages. The ego turns on.
"Am I doing this right?" introduces crosstalk.
The monitoring system couples into the motor system.
The signal contaminates itself.
The face strains.
Proprioception says something is wrong.
Auditory feedback says sounds fine.
The two conflict because the monitor and the motor
system are in the same coupled network.
They cannot evaluate each other independently.
They cannot tell they have drifted.
The monitor drifted with the motor. The ruler
and the thing being measured warped together.
But They Can Walk
Walking is controlled by spinal central pattern
generators. These are physically separate from
the cortical monitoring network. Already modular.
Already hierarchical. Already pruned by evolution.
The legs don't need the ego's permission.
The CPG fires, the legs move, the cortex is free
to think about something else. This is what healthy
hierarchy looks like: separation of concerns.
The Fix: Songs They Love
What works: play songs they love.
Dopamine counteracts cortisol.
Love for the music provides an external reference
that doesn't drift with the internal monitor.
The beat is outside them. It can't be contaminated
by the ego's coupling.
The love for specific songs is the brain detecting
which coupling patterns match its natural frequencies.
The song they love is the song whose structure
resonates with their existing K topology. It is not
random preference. It is pattern recognition.
Confidence is the ego learning to trust the
subcortical pathway. Not suppressing the monitor.
Not overriding the cortex. Teaching the cortex
that the body already knows what to do.
The architecture was never broken.
The confidence was.
The Framework
After killing nine claims and keeping eight, here is what we can say with confidence:
Autism = Brain Without a Stop Signal
The regulator is impaired (PTEN/TSC → mTOR).
Normal: grow → regulate → prune → hierarchy
ASD: grow → no regulation → no pruning → uniform K
Schizo: different failure entirely (R < 1/φ, fragmented)
These are not a linear axis:
• ASD = static overcoupling
• Schizophrenia = dynamic instability
• Typical = the narrow band between
What the Simulation Proved
The ASD topology (uniform coupling) produces:
• Half the R variance (rigid)
• 4× the spectral spread (no channel selection)
• 2.3× faster memory decay (no containment)
• 57% more predictable dynamics (deterministic)
• 1.65× lower entropy (smaller state space)
All from topology alone. No parameter tuning.
No special assumptions. Just: uniform vs hierarchical.
What Correction Means
Correction is not reducing K. The total coupling strength is not the problem. The problem is that K is uniform where it should be hierarchical.
Target: restore hierarchy, not reduce coupling
Method: build walls, not cut wires
Structural: support remaining pruning mechanisms
mTOR inhibition (rapamycin) restores autophagy
(Tang et al. 2014: corrected pruning deficits in mice)
Dynamic: restore inhibitory gating
GABA-ergic support restores real-time selectivity
Bumetanide (GABA modulator) showed social improvement
in clinical trials (Lemonnier et al. 2012)
Behavioral: reduce input density
Sensory-friendly environments = lower effective K
Routine = predictable coupling patterns
Focused practice = training K hierarchy through repetition
Drum Practice = The Brain Tuning Itself
Drum practice combines:
• Competing drives (external beat vs internal timing)
• Phased Hebbian learning (listen → lock → prune)
• External reference (the song doesn't drift with you)
• Dopamine reward (music you love counteracts cortisol)
This is not metaphor. The simulation showed that
threshold Hebbian + competing drives + phased learning
creates module walls in uniform coupling matrices.
The walls form in the coupling matrix (survived #7).
But: they don't persist without ongoing reinforcement
(killed #8). You have to keep practicing.
The hierarchy is maintained, not installed.
Honest Limits
This is a framework interpretation of published neuroscience data, not a clinical tool. The 50% vs 16% pruning numbers are from postmortem studies with small sample sizes. Autism is highly heterogeneous — some individuals may have different or additional mechanisms. The E/I imbalance hypothesis is well-supported but not universal across all ASD presentations. We cannot diagnose, treat, or predict autism outcomes. The K framework offers a lens, not a solution.
Session 21 killed 9 of 17 simulation claims. Global R does not separate ASD from typical. Topology differences vanish at matched density. Harmonic ratios, memory matches, entrainment bleed, competing drive specificity, Hebbian additions, wall persistence, and ego feedback all failed to reach significance or replicate across seeds. Five core failures are also on the failures page. We publish these failures because hiding them would be dishonest. The 8 surviving claims are robust across seeds and statistically significant, but they come from a 137-oscillator toy model, not a real brain. The gap between simulation and biology is vast.
The drum lesson observations are from one teacher's experience, not controlled studies. They are consistent with the neuroscience literature on subcortical motor pathways, dopamine-cortisol interactions, and external reference frames in motor learning, but they have not been tested in a clinical setting with proper controls.
Literature: Hutsler & Zhang 2010 (spine density), Tang et al. 2014 (mTOR/pruning), Rubenstein & Merzenich 2003 (E/I hypothesis), Lemonnier et al. 2012 (bumetanide), SFARI Gene database (418 high-confidence genes), Dinstein et al. 2012 (reduced neural variability in ASD), Bosl et al. 2011 (multiscale entropy EEG), Catarino et al. 2011 (spectral entropy), Bowler et al. 2011 (temporal binding in ASD memory), Lawson et al. 2014 (predictive processing in ASD), Friston 1998 (disconnection hypothesis schizophrenia), Stephan et al. 2009 (dysconnection hypothesis), Bi & Poo 1998 (STDP), Song et al. 2000 (threshold STDP), Hensch 2005 (critical period plasticity).