A control-theoretic account of a state that gets called “awakening” or “depersonalization” depending on who is watching.
Status. A measured dynamical model (a frozen five-state ODE) translated into the language of meditation-adverse-effects research. The model is measured; the mapping onto contemplative practice is an interpretation it motivates but does not license. Citations are footnoted.
The paper is published: Marshall, J. T. (2026). Self-Application as the Common Source of Benefit and Failure: A Control-Theoretic Model of Self-Regulation. Zenodo. doi:10.5281/zenodo.22943641 — every measurement in this post is traceable to it.
Record: https://zenodo.org/records/22943641
The field that studies meditation-related harm has a well-documented problem: the harm is real, it is common, and it is systematically filtered out of the record.
The numbers are not fringe findings:
Read those together and a shape appears. The phenomenology is common. The label is rare. The state is often experienced as good. And the instrument of record cannot reliably see it.
This post is an attempt to explain that shape — not “meditation is bad,” but why a failure mode this common stays this invisible.
Two documented observations are hard to reconcile without a mechanism:
Under-reporting is usually explained as stigma, or as teachers not asking. Both are real. But there may be a stronger, structural reason: the part of the system that would notice the failure is the part that survives it. If that is true, the under-reporting is not a communication problem to be fixed with better questionnaires. It is a property of the failure.
The rest of this post is a model that makes that precise.
The model maps onto the well-known triple-network picture — with one crucial addition: the three networks are not three modules. Two are generators and one is the regulator between them.
| neuroscience | model | what it does |
|---|---|---|
| DMN (default-mode) | G — generator of self-content |
produces the narrative self, the sense that there is a someone here |
| CEN (central-executive) | D — demand / reduction |
executive checking, reasoning, “working on” content |
| SN (salience) | the loop (a, S, g) |
attention (a), setpoint/context depth (S),
adaptive gain (g) — it switches and damps the other
two |
(This mapping is the interpretation layer — a hypothesis from the source it derives from, not a measurement. The model’s measured content is the dynamics, not the brain labels.)
A healthy settled system sits at G* = 0.886 —
self-content strong, the reduction demand D doing its job,
the cannibalization switch off (c = 0).
The “self” here is a level — how much self-content the
generator is producing — a control variable, not a thing.
Turn attention inward and hold it. In the model that is one parameter
— a, attention, drives to 1 — and it acts on the generator
in three ways inside dG/dt:
β_G·(1−a)·G·(1−G) —
self-content grows when attention is outward;−α_G·a·G — self-content
is consumed when attention is inward;−γ_G·G that
exists even at rest.Hold attention inward and you are subtracting from G
while the growth term starves. G falls. That is the whole
trick: watch the self and it thins.
But the thinning is not the dangerous part. Falling G
raises the error E = D − G — the reducer
still demands, the generator produces less — and the model has a switch
on that gap:
c = σ · max(0, tanh((E − Θ_eff) / w)), Θ_eff = Θ · S / S_restBelow threshold c = 0 and the healthy regime costs
exactly zero (measured:
max|ΔG| = 0.00e+00). Cross Θ_eff and
c ignites — and c feeds back into
attention:
da/dt = [ κ_in·(a_hold + χ·c)·(1−a) − κ_ext·u_ext·a − ρ_a·a ] / τ_aThat χ·c term is the trap. The failure captures
the attention that drives it. Inward focus lowers
G; low G arms c; c
pulls attention further inward; further inward drops
G more. A positive feedback loop, one direction, no brake
built in.
One clarification, because the intuition runs the other way
and it is easy to reverse: the thing that runs away is the
capture, not the self. There is no generator in
overdrive, no self running hot — that is the folk reading, and the model
does the opposite. Attention is held in place by the failure it feeds,
while the generator is consumed: G falls from
0.886 to 0.049 and stays there. This is a self switched off and
held off, unable to notice — not a self amplified. (The
distinction matters for what follows: the intuitive picture predicts a
runaway upward, and would look for overactivity. The measured
one predicts absence — and, per §6, an absent self is exactly what no
instrument is watching for. §8 collects this and three sibling
misreadings.)
Run the canonical episode — inward drive a_hold = 0.9
held past a duration threshold:
| quantity | healthy | after the transition |
|---|---|---|
G (self-content) |
0.886 | 0.049 |
c (the switch) |
0.000 | 1.000 |
E (error) |
−0.340 | +0.525 |
a (attention) |
outward | 0.882, inward, held |
And the transition is not gradual. Across a full sweep of the dose axes the model gives 214 healthy outcomes, 98 collapsed outcomes, and zero in between. The threshold is sharp and bisectable: a duration of 66.3 time-units at that intensity, an intensity of 0.5992, an affect amplitude of 0.1859.
This is the structural reason the approach is invisible from the inside: the observer is inside the loop being observed, and the loop is monotone. There is no signal announcing the edge, because the thing that would read the signal is the thing being consumed.
Here the model makes a precise, falsifiable claim about the kind of event this is — and the answer is not what a “spiritual emergency” account would predict.
Ordinary tipping points slow down as they approach (recovery times lengthen, variance rises). That is the basis of critical slowing down as an early-warning signal. This transition has none of it. An exact audit of the model found:
[−0.0015, −0.0055, −0.0088, −0.0427, −0.598], every mode
bounded away from zero;−0.0015 across five decades of distance to the
edge (spread: exactly 0) — recovery time is 667 time-units
whether you are far away or 2×10⁻⁶ from the cliff;What it actually is: a border-collision fold. The
healthy branch does not wobble and tip. It runs straight into the
switching manifold and is annihilated on contact — the
last healthy solution sits at ε_c = 0.265192279 and dies
exactly on E = Θ_eff (residual
10⁻¹³).
The practical consequence is a measurement consequence. The warning signal is not time, it is tolerance: what shrinks near the edge is the basin of attraction — the size of perturbation the system can absorb — tracking distance-to-edge almost exactly (slope ≈ 1). A system in this regime does not slow down before it fails; it gets fragile. Which means a protocol that measures only settling time will miss the approach entirely. You have to measure the size of the disturbance the system tolerates.
Two more measured facts, and together they are the mechanism for §2’s puzzle.
First, there is no second fold. The collapsed branch
has no mirror edge: traced in decreasing drive — including into the
unphysical negative region where drive becomes withdrawal — it simply
exists, all the way down. The hysteresis window is
[0, ε_c) — unbounded below. The state
persists at every physical drive below the fold, including
zero. You cannot exit by stopping the practice. (This
is the model’s version of why “just take a break” often does not work,
and why recovery in the clinical case is circumstance-linked rather than
self-initiated.) In the model, rescue is a hard floor pinned above the
collapsed error level — and it works, but only at or above
0.4795, essentially the state’s own error. Anything less
does nothing.
Second, reasoning survives and the self does not. At the collapsed fixed point:
D = 0.5455 ← exactly its healthy baseline
G = 0.049 ← annihilated from 0.886The executive capacity — the part that reasons, checks, and reports — is untouched. What collapsed is the generator of the narrative self. This is why the state reads as vast calm rather than distress: nothing is wrong by the reasoning, because the reasoning is fine. There is simply no one left for it to be about.
And this is the answer to §2. If the reporting
faculty (D) sits at baseline while the tracked faculty
(G) is annihilated, then:
a-axis. The sensor is destroyed by the failure it exists to
detect.That is the structural reason the under-reporting is not merely a questionnaire problem. The instrument of record is in the failing subsystem’s surviving neighbour, and the act of checking feeds the loop. The fix is not to monitor harder. It is to relocate the measurement to a channel that survives and cannot be moved by the loop — the body, the behavior, an external observer.
The picture most people carry of meditation “dark nights” is effortful — clinging, terror, white-knuckling. The model says the endpoint is the opposite. Tracking the trajectory through the model’s own plane (outward / relaxed-inward / effortful-inward):
a = 0.8824, D back at baseline —
with the switch armed for the rest of the run
(measured: 1318.70 time-units, c held at 1).So the endpoint is an effortless, contentless, locked openness. That is precisely why it is misappraised as the goal — and, per §1, why the same phenomenology lands as “more welcome, pleasant, and spiritually meaningful” in one population and as a disorder in another. The state that most needs care is the state that feels like the destination, and which no longer has the faculty that would ask.
The results above are counter-intuitive in one consistent direction, and readers — including this author — reach for the folk reading every time. It is worth naming the four, because each one changes what you would look for, and one of them changes what you would do.
| the intuitive reading | what the model actually shows | what changes |
|---|---|---|
| “The self runs hot, then collapses” — DMN overactivity, then annihilation | G reaches 0.8848 at the moment inward attention
begins, and never rises again; the decline starts immediately
(dG/dt = −0.0044/t.u.). The switch arms at t = 167.5, when
G is already down to 0.14. There is no
overactive phase to find |
the runaway is the capture, not the self — a search for hyperactivation finds nothing, because the predicted pattern is normal, then down |
| “Introspection is a resource that gets depleted” — you run out of fuel | Release the inward drive at t = 200 and give it 1000
time-units of nothing: G moves 0.0514 → 0.0486
(−0.0028 — slightly worse). Attention stays at 0.88 and
the switch at 1.0 throughout. It never refills |
overharvest, not a battery. A depleted tank refills when you stop draining it; this does not, because the draining never stops. (Note this is not ego-depletion theory, which is contested and makes the opposite claim about recharge) |
| “The reducer runs hard and eats the self” — CEN overdrive consuming the DMN | Inward attention alone collapses G
(0.885 → 0.0487) with D never leaving its equilibrium
(0.500 → 0.545, no spike). Conversely, raising D alone — an
affect pulse to 0.676, a 35% spike — collapses
nothing (G_end 0.8855, switch never arms) |
the reducer runs at its normal rate on a shrinking
substrate. D at baseline while G is
annihilated is the D/G dissociation, and it is the measurement
the sensor-placement result rests on. No CEN hyperactivity is
predicted |
| “The salience network loses the ability to switch” — the switch breaks | Nothing breaks. Two ordinary feedbacks latch it:
c holds attention inward (χ·c), and the
setpoint drains (S 0.865 → 0.167), so the arming threshold
Θ_eff falls 79% (0.64 → 0.134) — the bar
to stay on drops almost to nothing. And the capability survives intact:
applying external input post-collapse at u_ext = 0.8 drives
attention fully outward, restores G to 0.885, and disarms
the switch |
captured, not broken — and the cost is asymmetric:
u_ext = 0.3 prevents collapse if present during
the episode, but the same 0.3 does not rescue it
afterwards. Rescue costs more than prevention |
The first three are one family, and it is the family worth naming: the folk reading of a catastrophic loss is always something ran too hard. A self running hot. A fuel burned empty. A reasoner eating its host. In every case the model says the opposite — something stopped, and nothing came to stop the stopping. Generator annihilated, not amplified. Attention latched, not exhausted. Reducer unopposed, not overdriven.
And the fourth is the mechanism that makes the other three feel true. It is also the answer to “so what is actually failing?” — nothing. The exits are all still there; they just cost more than they did. That is why “just stop” and “just rest” don’t work, and why the coercion post’s malignant holder is so effective: the holder is standing at the one remaining exit, and the price of using it has gone up.
There is a line nearly everyone knows — do not go gentle into that good night — and this mechanism gives it a reading it rarely gets. It is normally heard as an instruction to the person facing the end: resist, hold on, fight. But Thomas wrote it to his dying father, who could not answer it. The raging in that poem is not the dying man’s. It is the demand of the living, aimed at someone who has stopped being able to supply it for himself. That is the tragedy, and it is structural.
The collapse is the same shape. The state does not rage — it goes gentle: an effortless, contentless, locked openness with the switch held on (§7), felt as peace, misappraised as arrival. The faculty that could feel the urgency, object, or ask for help is exactly the one sitting at baseline, seeing nothing wrong (§6). So “do not go gentle” cannot be addressed to the person in the state. It is addressed to whoever is standing outside it — and everything that follows is what that address obliges them to do.
None of the above is an argument against contemplative practice. It is an argument for informed consent, measurement that can see, and support that survives the state — the same provisions the traditions encoded and the adverse-effects field is now rebuilding.
The traditions are not the villain here; they are the precedent. They mapped this territory over centuries, and they built guardrails into the delivery. Three of them, and they are a stability spec, not mysticism:
| supplied beforehand | model term | what it prevents |
|---|---|---|
| A frame before, not after | — (meaning inherited, not built in-collapse) | having to derive a model of what is happening from inside the failure |
| Community present | external coupling, u_ext |
the coupling being improvised, alone, from the depersonalized state |
| Pace metered | slow drive ramp vs. the 66-unit threshold | the load arriving faster than adaptation |
The model says why these matter structurally. u_ext is
literally the outward pull that prevents capture; withdrawing from the
world to practise removes the damper. And the traditions’ sharpest
safety instruction — don’t kill the witness — was arrived at
empirically: it exists because the reports accumulated, because
practitioners in the lineages did the thing, got the loop, and reported
back. That is a warning system built from exactly the under-reported
cases §1 describes.
What follows, concretely:
66.3 time-units at that
intensity). The visible warning is not slowing but shrinking
tolerance, so the useful monitoring is perturbation tolerance, not
a settling time.The guard on that third point, because “regardless of consent” is the one line here that could be read as license. It means one specific thing, and it is not a general warrant. The decision is made in advance, while the person can still make it — because the state cannot initiate its own exit, the choice has to be made before arrival, and honouring it afterwards is what an advance directive is for. So the provision is not authority over someone; it is their own authority, exercised at the only point it is exercisable. It presupposes all three of a state recognised as severe, an existing relationship, and a holder who knows the signs — and the input is the outward pull itself (presence, contact, ordinary life), not a course of action substituted for theirs. Without those, the sentence reads as license. The mechanism licenses no one: it says only that the exit cannot come from inside the state.
The core is a deterministic five-state ODE. Every reference to brains or practice is an interpretation of the dynamics, not a measurement of them: the mapping is ordinal, unvalidated, and built from a single case, and four of five prior mappings from this model failed. Nothing here is a clinical claim.
The claim itself is narrow and falsifiable. The transition is a border-collision fold with no critical slowing down — recovery time constant to the edge, only tolerance shrinking — and recovery does not self-initiate. So the decisive test is specific: measure the perturbation the system tolerates, not just how fast it settles. Progressive slowing with a stable basin would falsify this account of the event.
That is the offer — a mechanism for the reporting gap, and one measurement to check it. The gap is the part that would matter if it holds; the fold is the part that could be wrong.
And one limit on its use, because a description of a mechanism is not a finding about a person. This is a model, not an instrument: it has no individual-level test, and the one population-level measurement that could give a signal has not been run. So it is not a way to assess anyone’s state, capacity to choose, or vulnerability — and in particular it is not a warrant to act on someone’s behalf against what they say they want. That inference — this person’s agency has been overborne, therefore acting for them is rescue — is the exact step that turned a body of careful thought-reform research into a movement that abducted adults. A description of an environment is not a finding about a mind, and the distance between those two is the whole of it. If you find this cited to override someone’s account of their own life, it is being misused — including by its author.
On the mechanism’s novelty, stated plainly: the general form — that measurement costs control performance — is classical dual control (Feldbaum 1960/1965). The specific form here — that the sensing channel coincides with the failure channel, so the fix is relocation rather than reduction — is the part we could not find stated elsewhere, with the caveat that sensor-placement and observability are mature control topics and a focused search is outstanding. The duality of self-application is Hofstadter’s strange loop, not a new observation. The contribution is the quantification.
Lindahl, J. R., et al. (2017). “The Varieties of Contemplative Experience: A Mixed-Methods Study of Meditation-Related Challenges in Western Buddhists.” PLoS ONE 12(5): e0176239. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0176239 (the study’s own framing: these experiences are “typically underreported”).↩︎
Britton, W. B., et al. (2021). “Defining and Measuring Meditation-Related Adverse Effects in Mindfulness-Based Programs.” Clinical Psychological Science. https://journals.sagepub.com/doi/abs/10.1177/2167702621996340 — n = 96; adverse effects 58%, negative impact on functioning 37%; lasting bad effects 6–14%, “associated with signs of dysregulated arousal (hyperarousal and dissociation).”↩︎
The Atlantic (Sept 2026), “The Dark Side of Mindfulness” — profile of Willoughby Britton; adverse effects lasting more than a month at roughly one in ten who have meditated, including after a single practice. https://www.theatlantic.com/ideas/2026/09/meditation-willoughby-britton-downsides/688474↩︎
Pons, E., et al. (2026). “A cross-sectional survey on depersonalization/derealization and meditation-induced alterations of the self.” Scientific Reports 16:14673. https://www.nature.com/articles/s41598-026-51014-y — N = 121 (60 meditation-triggered, 61 not); 61.7% of the meditation-triggered group above the CDS-70 cutoff; only 4 with any DPDR diagnosis; the states rated “more welcome, pleasant, and spiritually meaningful.”↩︎
Leavitt, F. (1999). “Dissociative Experiences Scale Taxon and Measurement of Dissociative Pathology.” J. Clin. Psychol. Med. Settings — the DES-T dissociative taxon false-positives at 54% in non-clinical samples (13% for the full DES). https://link.springer.com/article/10.1023/A:1026275916184↩︎
Pauly, L. et al. (2021). “Prevalence, predictors and types of unpleasant and adverse effects of meditation in regular meditators: international cross-sectional study.” BJPsych Open — N = 1,370 regular meditators; retreat attendance OR 1.89 (95% CI 1.43–2.48, p = 0.000), adjusted OR 1.94 (1.49–2.60). Replicates Schlosser, M. et al. (2019), PLoS ONE (25.4% unwanted effects). https://pmc.ncbi.nlm.nih.gov/articles/PMC8693904↩︎
“The Teacher Matters: The Role and Impact of Meditation Teachers in the Trajectories of Western Buddhist Meditators Experiencing Meditation-Related Challenges” (2025), Contemporary Buddhism, from the Varieties of Contemplative Experience dataset (68 practitioners, 33 experts) — 97% of practitioners and 97% of experts mentioned student–teacher relationships as an influencing factor “impacting meditation-related challenges, including their onset”; 71% were working with a teacher at onset; teacher support rated −3 (very harmful) to +3 (very helpful). https://www.tandfonline.com/doi/full/10.1080/14639947.2025.2485677↩︎
Sierra, M., Senior, C., Dalton, J., et al. (2002). Reduced skin-conductance responses to unpleasant pictures in depersonalization disorder — i.e. no autonomic recovery to baseline. https://pubmed.ncbi.nlm.nih.gov/12215083/↩︎
Pons, E., et al. (2026). “A cross-sectional survey on depersonalization/derealization and meditation-induced alterations of the self.” Scientific Reports 16:14673. https://www.nature.com/articles/s41598-026-51014-y — N = 121 (60 meditation-triggered, 61 not); 61.7% of the meditation-triggered group above the CDS-70 cutoff; only 4 with any DPDR diagnosis; the states rated “more welcome, pleasant, and spiritually meaningful.”↩︎