A number of conditions treated as psychiatric are, in some fraction of the people who have them, not psychiatric. They are what a person looks like when cortisol signalling is too low, or arriving with the wrong timing. What gets called depression, ADHD, seasonal depression, burnout, the exhausted end of PTSD, the collapse that arrives after years of OCD — all of these are diagnosed by symptom pattern, which means nothing in the diagnostic process ever asks whether the cause might be endocrine. The question is not answered in the negative. It is not put.
A note on how I am going to write about this, because it affects every sentence. I do not think the psychiatric categories are real things. ADHD, PTSD, bipolar disorder and the rest are lists of complaints that tend to co-occur, arrived at by committee and vote rather than by discovery, and the field's own leadership has said so: the diagnoses rest on "a consensus about clusters of clinical symptoms, not any objective laboratory measure" (Insel, 2013). So I will use the names, because you need to be able to find yourself in this, but I will use them as labels rather than as diseases — "what gets called ADHD" rather than "ADHD." It reads more awkwardly and it is deliberate. Writing as though the categories are entities concedes the assumption this page exists to argue with.
That it is worth asking is not speculative. Medicine runs the experiment by accident thousands of times a year: give people corticosteroids and a dose-dependent fraction develop mania, psychosis, or depression, and taper them and it mostly resolves (Warrington & Bostwick, 2006). Glucocorticoid signalling can manufacture states clinically indistinguishable from primary psychiatric illness, and take them away again. It can do that because cortisol is not a stress hormone with one narrow job — it gates glutamate release, the excitatory signal nearly every other system needs in order to fire, and it regulates the rate-limiting enzyme for dopamine synthesis. A shortfall therefore does not produce one clean symptom. It produces a scatter of unrelated-looking complaints across every system downstream, which is an excellent way to stay invisible to any single specialist.
Low cortisol in these conditions is well documented. What is wrong is the standard explanation for it, which blames unusually strong negative feedback and so makes receptor sensitivity the cause. That inverts the arrow. Receptors growing more sensitive is the ordinary response to scarce ligand, so enhanced feedback is what a low-cortisol state looks like from outside rather than what produced it. The likelier account is reduced drive from the top of the axis — the hypothalamus commanding less — which then holds itself in place in two ways. There is a behavioural loop, in which low cortisol means less movement and less light, which means still less drive. And there is a plasticity trap, which is the more interesting one: remodelling a synaptic arrangement requires NMDA signalling, NMDA signalling requires glutamate, and glutamate is exactly what is short. The system is missing the signal it would need in order to change state.
All of which would be cheap to check, except that the usual test is the wrong one. A single 8am serum cortisol misses this three ways, the worst being that the test demands the exact behaviour whose absence is the symptom: a barely-functioning person hauling themselves to an early appointment has mounted a stress response to get there, and that lifts the very number the appointment exists to measure. DHEAS is the better instrument — same adrenal drive, long half-life, no meaningful daily rhythm, so a single draw is interpretable and one hard morning cannot spoil it.
And still nobody looks, because no specialty owns the question. Endocrinology asks about diseased glands against thresholds and returns a normal result. Psychiatry sorts by symptom cluster and has no instrument for anything upstream of one. Functional medicine fills the vacuum with "adrenal fatigue," which is false, and which has made the legitimate version of the question sound like quackery. The result is people disabled for years while three sets of specialists each correctly conclude that it is not their department.
The rest of this page is that argument in detail, with the evidence, and with the guesses marked as guesses.
Everything below follows from one prior, so it should be stated up front rather than left implicit.
In any control system I have worked on, sustained overload almost never shows up as the actuator wearing out. It shows up as the controller changing what it commands. Actuators are simple, and simple things mostly either work or fail outright. Controllers are where the adaptation lives, and adaptation is where you get stuck states — a system that correctly reconfigured itself for a crisis and then never reconfigured back, because nothing in it was watching for the crisis to end.
Applied here: the adrenal cortex is an effector. It makes steroid on command. The paraventricular nucleus integrates limbic, brainstem, circadian, metabolic and immune input and decides what to command. So when someone is chronically under-cortisoled after years of load, "the gland got tired" and "the receptors wore out" are both proposing that a simple component degraded. "The controller turned its output down deliberately and then didn't turn it back up" proposes that nothing broke at all.
I find the second far more believable, and not only on taste. It requires no damage. It predicts reversibility, which matters because people do recover from this, sometimes years later, which is awkward for any account built on accumulated wear. And it does not need an epicycle to explain why the periphery looks fine — burnout patients do not have Addison's, their adrenals are unremarkable, and a damage model has to work to explain that. A stuck-controller model expects it.
This is also, I think, the defensible version of an idea that deserved its bad reputation. "Adrenal fatigue" was systematically reviewed and rejected (Cadegiani & Kater, 2016), and the rejection is correct: the adrenal is not fatigued. But it gets over-read as "therefore nothing is wrong with the axis," which does not follow. Nothing here requires the gland to have failed. The claim is that it is being asked for less.
Nor is the intuition original to me, which I take as reassuring rather than disappointing. It is most of the way to allostasis: McEwen's taxonomy of allostatic load already names "failure to shut off allostatic activity after stress" and "inadequate response of allostatic systems" as distinct forms (McEwen, 1998). That is a controller stuck in the wrong state, described in 1998. What I am adding is a guess about which controller and a way to check.
Two honest limits. This is a prior, not evidence — a rule about where to look first, not a finding. And complex neural tissue plainly can be damaged by stress; hippocampal dendritic remodeling is real and well documented. So the argument is not that the brain cannot break. It is that a regulated state change deserves to be the first hypothesis and component failure the second, and much of this literature reaches for them in the opposite order. Also, "the controller decided" is shorthand. Nothing decided anything. It is regulation and selection, not intent, and the language of engineering is doing some flattering work there that the biology has not earned.
A low reading and a low state are different things, and the difference carries all of the practical weight.
If cortisol were merely low, removing the load should let it return. Often it does not: the low state persists for years after the stressor is gone, which is the observation this whole page is trying to account for. That is the behaviour of a system with more than one stable operating point, where the low one holds itself and the way back is not the way in reversed.
Two things could make it self-holding, and they are not exclusive.
The behavioural loop. Low cortisol means low morning arousal, which means less light-seeking and less movement, which removes the two strongest inputs the circadian system takes, which weakens drive, which lowers cortisol further. There is no receptor pathology anywhere in that circle and nothing is damaged — it is ordinary couplings arranged so that the output feeds the input. It also runs in reverse, which is the only reason any of this is worth writing down.
The plasticity trap. This is the one I find more interesting, and it is the sharper version of what the rest of the page has been circling. Changing a synaptic arrangement requires plasticity, and plasticity at these synapses is NMDA-receptor dependent. NMDA receptors need glutamate. So a system short of glutamatergic signalling is short of precisely the signal it would need in order to remodel itself back. It is not just sitting in a bad state; it has lost the mechanism by which it would leave one. That explains permanence far better than accumulated wear does, and it predicts something wear does not: restoring signalling should restore the capacity to change, not merely raise the level while the input lasts.
The trap has a door that only opens inward. The half of that argument I had been missing is how the system got into the configuration in the first place, and it turns out to be the same mechanism running the other way. Stress does not merely act on these synapses; it opens a plasticity window at them. A single stressor primes glutamate synapses in the paraventricular nucleus so that they can undergo potentiation they otherwise could not, gated by CRH-dependent changes at postsynaptic NMDA receptors (Kuzmiski et al., 2010), and sustained stress produces frank structural reorganisation there — altered cell size, altered glutamatergic and GABAergic innervation (Herman et al., 2008). High adrenal output during a stressor means the substrate for plasticity is abundant at exactly the moment the system is rewiring itself.
So the configuration is easy to enter and hard to leave, for the same reason in both directions. Plasticity is available while the system is being driven hard, and unavailable once it has turned itself down. That is not ordinary hysteresis, where the return path is merely different from the outward one. It is a ratchet: the gate is held open by precisely the conditions that are about to vanish, and it closes behind you. Which would explain a pattern people report and nobody explains — that they can name the year this happened to them, and cannot name any year it reversed.
Why the exit matters more than being right. There is a well-developed competing account in which the trap is real but sits at the pituitary: include glucocorticoid receptor expression in a model of the axis and it becomes bistable, with a high-GR, low-cortisol state that is stable (Gupta et al., 2007). I have already said why I think the feedback reading inverts cause and effect. But there is a second objection that does not depend on being right about that. Gupta's trapped variable is pituitary GR, which has no natural, controllable input — nothing you or a clinician can push on. That trap is sealed. A drive-side trap puts the stuck variable somewhere with several continuous inputs that can actually be moved: light, movement, load, and possibly a small amount of cortisol. Both accounts describe a trap; only one of them implies anything to do about it. That asymmetry is a reason to test the drive-side account first. It is not a reason to believe it, and I want to be careful not to let usefulness masquerade as evidence.
The shape of recovery discriminates them. A sealed trap predicts a flat course and then a discrete jump when something finally crosses the barrier. A pushable one predicts graded, compounding improvement — a slow climb that accelerates, because the loops that locked the system down run forward once they turn over. Those two trajectories are distinguishable in a single person across months, which makes it among the cheapest tests on this page.
Everything above assumes the PVN turned its output down and has not turned it back up. It is worth asking what it consults when it decides, because the answer is not one number and the machinery has consequences.
Cortisol, on two tiers rather than one. There is no single linear sensor. Two receptors with different affinities sample the same signal: the mineralocorticoid receptor, roughly tenfold higher affinity, substantially occupied at basal levels and supplying tonic restraint, and the glucocorticoid receptor, lower affinity, filling only at stress peaks and the circadian peak. On top of that there are two timescales — slow genomic feedback over hours, and a fast non-genomic arm working in seconds to minutes. So the question "is suppression linear or does it only kick in above a threshold" has the answer: neither exactly, it is a baseline holding term plus a separate surge-handling term, sampled by two sensors and acted on at two speeds.
The fast arm operates on precisely the synapses this page is about. Glucocorticoid binds membrane receptors on CRH neurons, which triggers endocannabinoid synthesis; the endocannabinoid then travels backwards across the synapse to CB1 receptors on the incoming terminals, where it suppresses presynaptic glutamate release (Di et al., 2003). Fast feedback does not silence the cell directly. It reduces the glutamatergic drive arriving at it — the same variable Salter measured as reduced excitatory synaptic gain, and the same variable the entire plasticity argument turns on. The feedback mechanism and the lesion act at the same place, which is a coincidence worth staring at, and one with an uncomfortable implication for anyone supplying cortisol from outside.
Cortisol is only the feedback term. Something else issues the command. And it is not arranged the way most people picture it. Prefrontal cortex, hippocampus and amygdala have essentially no direct projections to the PVN. They act through GABAergic relays — the bed nucleus of the stria terminalis, peri-PVN neurons, the lateral septum. Hippocampus and medial prefrontal cortex impose inhibition through those relays; the amygdala drives the axis by disinhibiting them (Herman, 2020). Limbic control of this nucleus is exercised entirely by modulating GABAergic tone, which is the same knob everything else on this page keeps arriving at.
Two roads in, and they are not the same road. The literature separates systemic or physiological challenges — exercise, metabolic demand, homeostatic threat, routed through brainstem pathways — from psychogenic challenges routed through the limbic relays. Different circuits, converging on one nucleus.
That distinction converts a vague prescription into a specific one. If the goal is to drive this axis without reproducing the thing that broke it, the target is demand arriving as physiological load rather than as danger. A high-output, low-threat life is not a compromise between two goods; it is an instruction about which input to use. Exercise is the cleanest available instance — real demand, no threat content, and it generates BDNF while it is at it.
It is also the closest thing to an answer I have for what the PVN would be remodelling toward. Not "more cortisol." The capacity to support a load that arrives as work rather than as fear.
Ketamine is the clearest working example of forcing a plasticity window open from outside, which makes it the obvious thing to compare any of this against.
Its mechanism is not what the label suggests. Ketamine blocks NMDA receptors, but the blockade is the trigger rather than the treatment. Subanaesthetic doses preferentially block NMDA receptors on GABAergic interneurons — tonically active cells, so more of their channels are open to an open-channel blocker — which disinhibits the pyramidal neurons those cells were restraining and produces a burst of glutamate. That glutamate acts on AMPA receptors, and blocking AMPA abolishes the antidepressant effect entirely, which is how we know which arm is doing the work. AMPA activation then drives BDNF, TrkB, mTOR, and the synthesis of new spines (Li et al., 2010). An NMDA antagonist that functions, at circuit level, as a glutamate-raising drug.
The six-infusion schedule follows from what happens next. Spines born in a plasticity window start weak, and they survive only if they are used; otherwise they are pruned within days. Repeated windows let consolidation outpace elimination. Worth noting that the schedule was arrived at empirically — repeated dosing was seen to extend response — and the spine-stabilisation account was fitted to it afterwards. I think the account is right. It is still a story told about a number that came from outcome data.
Which sets up the question this page cares about. Ketamine's synaptogenesis is documented in prefrontal cortex and hippocampus. Whether it durably remodels the PVN is, as far as I can find, simply not established. And the flat claim that ketamine misses the PVN is wrong in an interesting way: subanaesthetic ketamine roughly doubles circulating cortisol within an hour, so it plainly drives the axis (Birnie et al., 2022). It reaches the circuit. What is unknown is whether anything is rebuilt there, and there is a complication in the way of finding out, since that same cortisol surge is a candidate confounder for ketamine's antidepressant mechanism generally.
Does the surge reach the PVN? That is the whole question, and it is worth being careful about what is and is not known. The remodelling cascade — glutamate, AMPA, BDNF, TrkB, mTOR, new spines — is not specific to cortex. PVN neurons have all of those parts. So if a glutamate surge occurs there, it should build synapses there, by the same route it does anywhere else. Nothing about the hypothalamus exempts it.
What the surge requires locally is the disinhibition motif: tonically active GABAergic cells, bearing NMDA receptors, restraining the neurons of interest. The PVN is under heavy GABAergic tone, so the arrangement is at least plausible. And there is a piece of indirect evidence in favour that is easy to skate past — if ketamine roughly doubles cortisol within the hour, then PVN output has gone up, which is what disinhibition of this nucleus would look like from outside. The acute endocrine effect everyone treats as a side effect is a readout of the mechanism arriving.
What is actually missing is that nobody appears to have measured spine density or synaptic reorganisation at the PVN after ketamine. That is an absence of data rather than a negative result, and the two should not be confused.
A second mechanism, pointing the same way. There is also a plasticity route specific to this structure. The priming that allows PVN glutamate synapses to potentiate after a stressor is CRH-dependent (Kuzmiski et al., 2010) — CRH depresses postsynaptic NMDA receptors, lifting a retrograde brake on glutamate release. That is short-term potentiation rather than structural growth, so it is not a gate on all plasticity here, and I do not want to overstate it. But it means CRH is permissive for at least one form of plasticity at these synapses, and since anything that drives a glutamate surge at the PVN also drives CRH, the two mechanisms would run together rather than compete.
It does have one uncomfortable implication for the rest of this page. If CRH is permissive, and exogenous cortisol suppresses CRH, then the intervention I am running may be working against that particular route even while it supplies substrate for the general one. That tension is taken up below.
The version of the ratchet that survives all this is the general one rather than a PVN-specific gate: a system short of glutamatergic signalling lacks what remodelling requires, and stress supplied it abundantly on the way in. That is enough to make the trap asymmetric without needing CRH to be a master switch.
The PVN's version might be disinhibition too, and it might be a steroid. Ketamine's trick is disinhibition — take the brake off by silencing the cells applying it. Ask what else could apply that trick to this structure and an answer falls out of things already established separately. DHEAS is a negative allosteric modulator at GABA-A receptors, acting at the benzodiazepine site and a second, lower-affinity site. The PVN sits under heavy GABAergic tone, from local hypothalamic and adjacent forebrain neurons projecting onto its CRH cells. And disinhibition through GABAergic relays is the native route by which upstream structures such as the amygdala drive this axis at all.
Put those three together and DHEAS looks like a candidate PVN disinhibitor: the same strategy ketamine uses in cortex, at a different receptor, applied to the structure ketamine may not be rebuilding. I want to be exact about the status of that. Each of the three premises is established. The conclusion is not — I can find no one who has shown DHEAS modulating GABAergic input at the PVN specifically. It is an inference, and it is the kind that turns out wrong often enough to be worth labelling. The concentration problem applies as well: the GABA-A antagonism is characterised in vitro, and whether oral DHEA reaches levels that matter at the PVN in a living person is unknown.
A data point, and its limits. I have had two full courses of ketamine. The first worked dramatically — it reorganised my default mode network and changed my life. The second, run after the drive collapse this page is about, did nothing at all. Same drug, same person, same protocol, different state.
The model predicts that ordering. Ketamine needs a glutamate surge, the surge comes from disinhibiting cells so they release glutamate, and if presynaptic glutamate is depleted then taking the brake off an engine with no fuel achieves very little. Signalling was intact the first time and was not the second. It is also possible that the difference was load rather than substrate — the first course happened during a high-demand life, so the window had something running through it — but those are not competing explanations so much as the same state difference described twice.
What this is not is a controlled comparison. Novelty and expectation differ enormously between a first course and a repeat, diminishing response to repeated ketamine is independently well known, and n is two. I include it because a within-subject reversal is more informative than the usual anecdote, not because it settles anything.
There is one consolation in the comparison. Ketamine opens a window in a clinic, in a recliner, and the plasticity peaks a day later wherever that happens to fall — the window and the activity are decoupled, which is a poor arrangement given that use is what stabilises a new synapse. Load has no such problem. The demand and the plasticity it licenses arrive together by construction.
The thing that got me into this is that cortisol is an input to a surprising amount of machinery, so a shortfall doesn't present as one tidy syndrome. It shows up as an assortment of unrelated-looking complaints, which is a good way for a deficit to stay invisible to any single specialist.
Two examples that matter for what follows. Glucocorticoids regulate glutamate release, and glutamate is the excitatory signal nearly every other neurotransmitter system needs in order to fire (Popoli et al., 2012). And glucocorticoids regulate tyrosine hydroxylase, the rate-limiting enzyme for dopamine synthesis — corticosterone raises TH expression in brainstem culture (Busceti et al., 2019), and the effect goes back to the older dexamethasone work (Oka et al., 1985).
That second one is worth sitting with, because it means the ordinary "ADHD is a dopamine problem" story and a cortisol story are not competitors. If you are chronically low on effective cortisol signaling, running low on dopamine is a downstream consequence, not a separate disease. The dopamine finding could be perfectly correct and still not be the place to intervene.
It is worth spelling out why a glutamate shortfall produces something qualitatively worse than low mood, because this is the part that makes the whole account matter. Glutamate is not one transmitter system among several. It is the excitatory input the others need in order to fire at all: the dopamine neurons of the midbrain, the serotonin neurons of the raphe, the noradrenergic cells of the locus coeruleus and the oxytocin neurons of the hypothalamus are all driven by glutamatergic input, and GABA is synthesised from glutamate directly. It sits underneath the entire cast rather than beside it.
So a shortfall does not turn one channel down. It turns all of them down at once, which is a different experience from ordinary anhedonia. Ordinary low mood is reward that fails to land. This is reward, salience, arousal and connection all under-signalling simultaneously — nothing lands, including the things that were never about pleasure. What that produces is flatness rather than sadness, and people describe it in those terms: not sad, just nothing, which is usually heard by a clinician as depression with poor insight.
It also predicts why single-transmitter drugs underperform in this group. Raising serotonin, or flooding the striatum with dopamine, acts on one output of a system whose shared input is short. You can push hard on a single channel and get less than the dose should buy, because the carrier signal underneath it is weak. That is not a reason to think the drugs do nothing. It is a reason to expect them to disappoint in a way that gets recorded as treatment resistance.
There is a timing consequence to this that I have not seen drawn out. TH is induced rather than switched on, and it decays slowly: induce it in rat adrenal medulla with cold stress and activity peaks around 24 hours, then falls away with a half-life of roughly three days, the enzyme protein itself turning over on a similar scale (Chuang et al., 1975). So a cortisol spike does not buy an hour of dopamine. It buys several days of raised synthesis capacity.
Which predicts something specific and, I think, recognisable: after a real stressor a person may be noticeably more functional for several days and then come down — not because the stress did them good, but because they are running on an induced enzyme pool that is still decaying. It would account for a pattern that otherwise looks perverse. Crisis making people sharper. Deadlines working. The collapse arriving in the week after the event rather than during it.
It is also the same structural move as the FKBP51 argument below: the phenotype has a time constant set by protein turnover, so it lags its input instead of tracking it. Anything sampled on the wrong timescale will look like noise.
Caveat: the TH literature is mixed rather than settled. Adrenalectomy-plus-replacement studies have not consistently shown the corresponding change in central dopamine neuron activity, so "cortisol is required for dopamine synthesis" overstates it. "Cortisol is one of the regulators of the enzyme, and the interaction is real but context-dependent" is about as far as the evidence goes.
That cortisol runs low in stress-related conditions is not my idea and not new. Yehuda's group has been reporting low urinary cortisol in people carrying a PTSD diagnosis, alongside exaggerated suppression on low-dose dexamethasone, since the late 1980s (Yehuda, 2002). The observations are not what I disagree with.
The standard reading of them is that the axis is being clamped from above: glucocorticoid receptors are unusually sensitive, so negative feedback is unusually strong, so cortisol settles low. Receptor sensitivity is the cause, low cortisol is the effect.
I think that runs the arrow the wrong way. Upregulating receptors when ligand is scarce is the most ordinary homeostatic response there is. If cortisol is chronically low for any reason at all, you should expect receptors to become more sensitive, and you should expect exaggerated suppression on a dexamethasone challenge. Enhanced feedback is simply what a low-cortisol state looks like from outside. Observing it tells you the state exists; it does not tell you what produced it.
So the standard account takes a predictable downstream consequence and promotes it to cause. It also has to posit a primary receptor abnormality to do it, where reading the arrow the other way requires nothing but a control system responding to reduced input the way control systems do. Between the two, the second is the one that needs no special pleading.
Which leaves the actual question open: something is holding cortisol down, and the enhanced feedback is a readout of that, not its source. The candidate I find most plausible is reduced drive at the top of the axis. Chronic stress dampens excitatory synaptic gain onto neurons of the paraventricular nucleus (Salter et al., 2018), and rats recovering from chronic variable stress show blunted ACTH and corticosterone responses for days after the stressor is gone (Ostrander et al., 2006). That is a drive problem. The axis is not being over-suppressed downstream; it is being commanded less.
This also matches the phenomenology better, at least mine. Exhaustion does not feel like signals arriving and failing to land. It feels like nothing is being sent.
The standard screen is a single serum cortisol at 8am, and for this question it is close to the worst instrument available.
It is one point on a curve, first of all, when the deficit is in the shape of the whole day. An all-day salivary profile shows a flattened or sagging curve that a single number cannot represent. Second, the reference range assumes a standard wake time: the morning rise is keyed to the sleep-wake transition, and diurnal cortisol activity tracks when a person actually woke (Edwards et al., 2001). If your sleep has shifted, which it usually has in this state, 8am is simply a different clock position for you than for the population the range was built from.
Third, and this is the one I think matters most: the test requires the exact behaviour whose absence is the symptom. To make an 8am blood draw, a low-functioning person has to set an alarm, get up early, travel, and be somewhere on time — a genuine acute stressor and a wholly atypical day for them. Someone who is at work at nine every day does the same thing as routine and at no cost. So the measurement conditions are systematically harsher for the sicker person, and the artifact runs in the direction that erases the finding: the effort of attending the appointment raises the number the appointment exists to measure. The worse off you are, the more the act of being measured hides what is being measured.
This is the strongest argument for the marker in the next section. DHEAS has a long half-life and no meaningful diurnal rhythm, so it integrates over weeks and cannot be spoiled by one hard morning. It is the readout that does not care how you got to the lab.
The useful consequence of all this is that the two accounts are distinguishable with a blood test, because cortisol is not the only thing the adrenal cortex makes under ACTH. DHEA and its sulfate DHEAS come off the same drive. So they dissociate informatively:
DHEAS is a better instrument than cortisol for this, which is the part I wish were better known. It has a long half-life and essentially no circadian rhythm, so a single draw means something — whereas a single cortisol reading is close to uninterpretable, being pulsatile, diurnal, and responsive to the experience of having blood taken. In secondary adrenal insufficiency, DHEAS was low in 80 of 84 untreated patients, and low more often than cortisol was (Yamaji et al., 1987). The marker of drive turned out to be more sensitive than the hormone everyone actually orders.
Two limits on how hard this can be pushed. Low DHEAS points upstream to ACTH, but it does not by itself separate hypothalamus from pituitary — PVN is where I would put my money, for the reasons above, but telling the two apart properly needs ACTH measured alongside and ideally a CRH stimulation test. And DHEAS falls steeply with age, so "low" has to mean low for your age rather than under the lab's floor, which is a distinction routine reference ranges are bad at.
People routinely talk about T/T carriers as though they are cortisol resistant the way someone is lactose intolerant — a fixed trait you have all day, every day. That cannot be right, and the reason is a clock.
FKBP51 is an induced protein with a short half-life: roughly 8 hours for isoform 1, 4 for isoform 2 (Martinelli et al., 2024). Cortisol drives its transcription, and in T-allele carriers each cortisol pulse produces more of it. But it burns off. If T/T produced standing resistance, the pool would ratchet upward across a lifetime of cortisol exposure and carriers would become progressively more resistant with age until they were non-functional. That is not what happens.
So the phenotype is dynamic, not static. What T/T buys you is a steeper induction curve on a pool that decays over hours — resistance that is a function of your recent cortisol history rather than your genotype alone. On an eight-hour window it looks like resistance. Across a life it looks like a thermostat with lag.
This predicts intraday texture rather than a constant trait, and the texture is testable: a stressful morning should produce an afternoon trough as the induced pool peaks, and then a second wind in the evening as it clears. If T/T carriers turn out to be unusually attached to afternoon naps and unusually alive at 10pm, that is not a personality quirk, it is a protein half-life. I have not found anyone who has looked, and diurnal FKBP5 work in humans would be a cheap way to check.
What gets called ADHD bundles two things that look like opposites: long stretches where nothing will engage at all, and episodes of total absorption that are hard to interrupt. A deficit model has to treat one of them as the anomaly, and it generally picks the second, filing sustained intense focus as a symptom.
I am not willing to write it that way. Something mattered to you and you engaged with it completely. That is not a malfunction — it is most of how anything good has ever got made. Calling it a symptom is a judgement about conformity wearing a clinical costume: it measures distance from the population average rather than distance from anything worth wanting. If a person gets upset about a problem and disappears into it for eleven hours, the honest description is that they cared and were able to act on it. The world could stand more of that, not less.
What is actually wrong is the part in between, and it is the part nobody names. If the functional range is compressed — activation has to clear a high threshold before anything engages — then below that threshold nothing recruits, and once something finally crosses it there is no graded middle to modulate with, so recruitment goes straight to the top. One range, two ends, and no dial. The complaint is not that the high end exists. It is that the intermediate settings are unreachable, so you cannot bring moderate effort to a moderately important thing, which is what most of a life is made of.
That changes what an intervention would be for. The target is restoring the middle, not suppressing the top — and it makes flattening a suspicious sort of success, since anything that narrows the range from above will score as improvement on a rating scale while removing the most valuable capacity the person had.
FKBP5 variants have been associated with the ADHD label and with diurnal cortisol in children (Isaksson et al., 2015; Kim & Jin, 2021), which is at least consistent, though these are small candidate-gene studies and should be read as such.
The dominant account of seasonal depression is about timing: circadian phase drifting out of alignment. I do not think that is wrong so much as incomplete, because it does not explain the people whose problem is not that they are shifted but that they are flat all day.
The alternative is that the light is arriving on schedule and registering at reduced amplitude. Bright light in the morning normally produces an immediate cortisol rise (Leproult et al., 2001). If cortisol is already low and glutamate signaling with it, the photic signal transduces weakly, the morning cortisol response is blunted, and cortisol stays low — which means the next morning's light registers just as weakly. It is a loop, and loops are why winter can lock in rather than simply tracking day length.
There is a concrete molecular reason this is possible, and it is stronger than the rest of the section. Light does not reach the clock as some abstract signal — the retinohypothalamic tract is glutamatergic, and photic entrainment of the suprachiasmatic nucleus runs predominantly through glutamate acting on NMDA receptors, to the point that activating those receptors directly produces light-like phase shifts in vivo and the phase-response curve for glutamate matches the one for light (Mintz et al., 1999). Light is a glutamate signal at the clock. So a glutamatergic deficit does not merely coexist with poor light response; it attenuates light at the first synapse it arrives at.
Framed this way it is a gain problem rather than a phase problem, and the two make different predictions. A phase problem should respond to correctly-timed light of ordinary intensity. A gain problem may not respond to ordinary intensity at all, because the input is being multiplied by a number that has gotten small — which would be one explanation for why light therapy works well for some people and does nothing for others.
DHEA probably belongs in this loop too, not only as the diagnostic marker it is above but as a contributor to the excitatory floor itself, so that low adrenal output reduces the gain twice over. Its positive modulation of NMDA receptors is real but weak, considerably weaker than pregnenolone sulfate (Johansson et al., 2010), so I would not put much load on that particular link — its negative modulation at GABA-A is the better-characterised action and probably the one that matters.
There is a natural experiment running constantly that I think deserves more weight than it gets. Give someone corticosteroids at dose and a meaningful fraction develop hypomania, mania, or frank psychosis — around 1.3% of patients at 40mg prednisone equivalent or below, 4.6% between 41 and 80mg, and 18.4% above 80mg (Warrington & Bostwick, 2006). It is dose-dependent. It mostly resolves on taper. Longer courses tend toward depression instead.
So glucocorticoid signaling can produce mood episodes that are clinically indistinguishable from primary psychiatric illness, and can take them away again, in a dose-responsive and reversible fashion. That is not a subtle correlational finding. It is something medicine does to people by accident thousands of times a year and writes up as an adverse effect.
Which raises an obvious question about the people who arrive at a clinic with the same presentation and no prescription. For any given patient, "is this one cortisol-driven?" has a real prior, and the bipolar label is the case I would most want checked. Nobody checks.
The reason nobody checks is structural rather than anybody's fault, and it is not a fringe complaint. Psychiatric diagnosis is descriptive by design: it sorts by symptom cluster, not by mechanism. When NIMH pulled its funding away from DSM categories in 2013, its director wrote that the manual's weakness was "its lack of validity," and that "unlike our definitions of ischemic heart disease, lymphoma, or AIDS, the DSM diagnoses are based on a consensus about clusters of clinical symptoms, not any objective laboratory measure" (Insel, 2013). That is the head of the largest funder of mental health research in the world saying the categories do not carve at mechanism.
The consequence is that the absence of evidence here is partly manufactured by the method. If the diagnostic system sorts by presentation and never asks what is upstream, then a cortisol-driven case and a non-cortisol-driven case land in the same bucket, receive the same treatment, and the difference between them never surfaces as anything except unexplained variance in who responds. Which is, more or less, what the outcome literature reports for nearly every psychiatric condition.
I want to be careful about how hard I lean on this, because "cortisol explains a lot of psychiatry" is one short step from "cortisol explains everything," and that is the genre this page is trying to climb out of. The claim is not that these conditions are cortisol disorders. It is that the question is answerable, cheap to ask, and largely unasked, so confident negatives are not available either.
What gets called OCD is where I would make the strongest guess, because the first half of the chain is already documented. Meta-analysis puts cortisol higher in people with the diagnosis than in controls with a moderate-to-large effect (d = 0.76; Sousa-Lima et al., 2019), alongside reports of elevated CRH in cerebrospinal fluid, raised early-morning plasma cortisol, and increased nocturnal ACTH. Whatever else OCD is, it is a sustained high-drive state on this axis.
Run that for years and the rest of this page predicts what happens: the controller does what controllers do under sustained overload, turns the drive down, and does not turn it back up. The person lands in exhaustion. That would make a meaningful share of "OCD patient who became treatment-resistant and exhausted in their thirties" a single trajectory rather than a comorbidity.
The part I find genuinely interesting is that the remodeling would work. Less drive means less of the arousal that feeds the obsessive loop, so the compulsions ease. The local objective is met. The global cost is a person who no longer has the activation to run their own life. A clinician seeing reduced compulsions alongside a flattened, low-functioning patient may be looking at a control system that solved exactly the problem it was given, at a price nothing in the loop was measuring.
That generalizes past OCD, and it is the reason I think it is worth writing down: any psychiatric improvement that arrives together with global flattening is worth reading as a possible trade rather than a partial recovery.
Status of this one: the first half is documented, the second half is a guess with nothing behind it but the argument above. It is testable, and cheaply — follow an OCD cohort and see whether symptom scores fall as exhaustion develops, with DHEAS alongside to say whether drive is dropping at the same time. As far as I can tell nobody has looked.
Everything on this page is sourced to peer-reviewed work. The PVN findings, the DHEAS discrimination, the glutamatergic light path, the steroid psychiatry — none of it is hidden or heterodox. Researchers have been doing their job and publishing it for decades. The failure is not that the knowledge does not exist. It is that nothing in clinical practice is arranged to act on it, and a person with this presentation falls through the space between three specialties, each of which fails them in a different and quite specific way.
Endocrinology asks a question this does not answer. The workup is built around gland pathology with thresholds: is it primary or secondary, is ACTH high or low, does the stimulation test pass or fail. Those are good questions and they catch the diseases they were designed to catch. But a functional central under-drive that never crosses a diagnostic cutoff is not a finding in that framework — it is a normal result. You are told your axis is fine, and by the only question that was asked, it is. The idea that the brain could be commanding less without any gland being diseased is not rejected so much as unasked, because there is no box for it on the form.
Psychiatry has no interest in mechanism, by construction. This is not a slur, it is the field's own account of itself: the categories sort by symptom cluster and were never meant to carve at cause. Its largest funder said so out loud, that the diagnoses rest on "a consensus about clusters of clinical symptoms, not any objective laboratory measure" (Insel, 2013). The consequence is a specialty that owns every symptom in this document and has no instrument for anything upstream of them. It does not measure the axis, it does not prescribe hydrocortisone, it does not consider DHEA, and it has no procedure for the question "what if this presentation has a cause." A patient asking it is not refused. They are simply speaking a language the appointment does not parse.
Functional medicine fills the vacuum with something false, and does the most damage doing it. "Adrenal fatigue" was reviewed and rejected on the evidence (Cadegiani & Kater, 2016), and the rejection is correct — the gland is not tired. The harm is not just that people are sold a wrong mechanism and supplements to match. It is that the wrong version now occupies the space where the real question would go, so that anyone raising central hypocortisolism in a clinical setting sounds like they have been reading the wrong websites. The quackery has made the legitimate question un-askable. Of the three failures this is the one I find hardest to forgive, because it was avoidable and because it discredits the patients rather than the practitioners.
There is one more piece of the structure, and it is the part I find genuinely funny. A doctor who read everything on this page, agreed with it, and acted on it would be exposing themselves — prescribing outside a diagnosis, departing from a standard of care that is defined by what peers customarily do rather than by what happens to be true. The sanction attaches to deviation, not to error. A clinician can be wrong in the ordinary, consensus way for an entire career and remain protected; be right in an unusual way once and they are not. I have no license, so there is nothing to take away from me, which is the only reason a page like this gets written from outside the system rather than inside it.
Which is not evidence of anything, and it would be cheap not to say so. Freedom from professional consequence is exactly the freedom the adrenal-fatigue industry enjoys, and it is perfectly symmetric: it lets me write this and it lets them write that. Being unsanctionable is not a form of being correct. What it does explain is the shape of the gap — how findings can sit in plain view for decades without reaching a clinic. Reading the literature is free. Acting on it is not.
None of this is a conspiracy and I am not claiming anyone acted badly. It is a structural gap: the diagnostic categories were drawn before the mechanism was understood, and nobody owns the space between them. But the effect on a person is not structural or abstract. It is years of being functionally disabled while three different specialists each correctly conclude that it is not their department.
This is what I would do if I had low cortisol, and I know that because it is what I am doing. I am not going to stage it as a case report or hedge it into meaninglessness. I have read the literature this page cites, worked through what it implies, and I am running the conclusion on myself. That is the basis. Weigh it accordingly, and weigh it against the alternative on offer, which is a specialty that does not measure the axis and a specialty that returns a normal result because it asked a different question.
What follows is the regimen, the reasoning, the three things I think are wrong with it, and the ways it can actually hurt you. The last two are not disclaimers bolted on for cover. They are the parts you would need in order to decide anything for yourself, which is the entire point of writing it down.
What it is. Oral hydrocortisone, a maximum of 2.5mg across a whole day, broken into very small amounts rather than taken as one or two doses, weighted toward the early part of the day. DHEA on roughly the same shape. Bloodwork periodically.
How, physically. There is no elegant apparatus for this and it is better to be explicit than to leave people picturing a pill organiser. I break open a capsule, tip the powder out onto a surface, and every hour or two wet a fingertip and touch it to the powder. That is the dose. It is not weighed and at that resolution it is not really weighable — a fingertip is however much sticks to a fingertip.
Two things follow from that and both are worth saying. The daily total is set by what was in the capsule and how much of it I get through, not by any per-dose figure, so anyone doing this should know the strength of what they opened, because that number is the one doing the work. And the imprecision cuts against me: it means my own regimen is not reproducible, not reportable, and harder to evaluate than it would be if I were measuring. I am describing what I do. I am not defending it as method.
For scale, because "low-dose hydrocortisone" covers a wide range: an average person produces on the order of 10mg of cortisol a day, and full replacement for adrenal insufficiency is 15 to 25mg. So 2.5mg is around a quarter of normal daily production and roughly a tenth of a replacement dose. The spreading-out is deliberate rather than fussy — if the target is a slow structural process rather than a fast symptomatic one, sustained low-level input is the matched shape, and a single pulse would be the wrong waveform for it.
What the DHEA is probably actually doing. I described the DHEA above as support for glutamatergic signalling, on the basis that DHEAS positively modulates NMDA receptors. That is true but weak, and I now think it is the smaller half of the story. The better-characterised action is the opposite sign at a different receptor: DHEAS is a negative modulator at GABA-A, and on the argument above that would make it a disinhibitor of a structure held down by GABAergic tone.
If that is right it reverses which of these two drugs is the interesting one. Hydrocortisone supplies substrate and tells the controller there is no error. DHEA supplies a little substrate and takes a brake off, without suppressing ACTH at all. One of them substitutes for drive and one of them raises it. That is a real difference and it points the same way as everything in the section below: if a taper is ever going to work, it runs through leaning on the arm that does not lie to the controller.
The two arms come apart in practice, and the way they come apart is informative. I ran DHEA on its own first, before I worked out the hourly dosing, and the effect was large. But it was specific: my brain worked and my body did not. Thinking came back, and physical energy for anything like exercise did not.
That is what the pharmacology predicts, which is the interesting part. DHEA is a neurosteroid and a sex-steroid precursor with essentially no glucocorticoid activity, so it can act on central receptors while leaving untouched the substantial peripheral job cortisol does. Glucocorticoids are permissive for vascular reactivity to catecholamines — they do not raise blood pressure alone but they are required for the vascular system to respond properly to the signals that do. They regulate PNMT, the enzyme converting noradrenaline to adrenaline. And they act at the tyrosine hydroxylase step to restore catecholamine synthesis when it has been run down by physical exertion, which is the same enzyme discussed above in its dopamine role. People with adrenal insufficiency measurably underperform on exercise testing for exactly these reasons.
A methodological note, since this page is otherwise sceptical about self-report. A domain-specific dissociation is harder for expectation to manufacture than a general improvement. "I feel better" is what belief produces. "Cognition restored, physical capacity absent, and the split falls precisely along the line between central neurosteroid action and peripheral glucocorticoid action" is a more particular result, and I did not know that division of labour when I noticed it. It is not proof. It is a better shape of evidence than the usual.
Which also means one of these is available and one is not. DHEA is sold over the counter in the United States, cheaply, without a prescription, which matters for a page like this because it means the argument is testable by the people it concerns rather than only by people with a cooperative prescriber. That is not universal — it is prescription-only or unavailable in the UK, Canada, Australia and much of the EU. And being a supplement rather than a drug, neither the label dose nor the contents are guaranteed.
Why the DHEA is split up. This has a separate rationale from the hydrocortisone and it is the part people get wrong. Endogenous DHEA is secreted on a diurnal curve driven by ACTH, roughly in step with cortisol — it arrives across the day rather than in one delivery. Swallow a whole tablet and you get a peak instead, most of which meets first-pass hepatic metabolism on the way through: sulfated to DHEAS, and pushed down androgenic routes, with the oral route reported to raise downstream DHT metabolites several-fold. What that buys is a pulse of sex steroids and a higher DHEAS number. Neither is the objective. The objective is free DHEA present continuously, which is the form that supports the signalling the rest of this page is about.
The comparison that makes the point is the same logic as an estrogen patch rather than a pill. Percutaneous DHEA holds all measured steroids at essentially constant serum concentrations across 24 hours, where oral dosing does not (Labrie et al., 2007). Route determines whether you get a plateau or a spike, and the plateau is what the body was doing on its own.
Three honest qualifications. Splitting an oral dose does not escape first pass — every fraction still goes through the liver, so this approximates a plateau rather than achieving one, and the consistent extension of my own argument is that transdermal would be better than split oral, which I have not tried. The same comparison study found no qualitative difference in which metabolites the two routes produce, so the strong version of the claim — that oral specifically manufactures androgens — is not clean; what is solid is the kinetics. And DHEAS is not waste: it is a circulating reservoir, retro-converted back to DHEA in tissue by steroid sulfatase. Raising it is not a failure. It is simply not the thing being aimed at, which matters because it is also the number you would otherwise be watching.
Why, on this page's theory — and it is not replacement. The goal is not to supply cortisol indefinitely. It is to raise signalling enough that the inputs which normally drive the system can land again, and then to let load do the work. Concretely: enough cortisol to support glutamatergic signalling; enough glutamate that light registers at the clock, where light arrives as glutamate, and that NMDA-dependent plasticity is possible at all; and then enough capacity to train hard and get through a full day, which is itself the demand signal telling the axis what it is being asked to support. On the plasticity-trap reading, this is the perturbation that restores the system's ability to remodel, rather than a substitute for its output.
Which sets the real test, and I cannot yet report on it: whether it can be withdrawn once load is carrying the state. If the model is right, it should eventually come out. If it turns out I cannot stop, the model was wrong about which thing I was fixing.
This is the strongest objection to what I am doing and I have not answered it.
There are two separate problems and I have only addressed one. The substrate problem was that without glutamatergic signalling no remodelling is possible in any direction. The setpoint problem is: granted that remodelling is now possible, what tells it which way to go? Cortisol is the feedback variable. Supplying it from outside is precisely how you inform the controller that there is no error to correct.
There is a narrower version that bites harder. CRH is permissive for at least one form of plasticity at PVN glutamate synapses, and exogenous cortisol suppresses CRH — so hydrocortisone may work against that particular route while still supplying substrate for the general one. That is not the whole story, since the main remodelling cascade does not require CRH. But a treatment that partly closes one of the doors it is meant to open is worth stating plainly rather than hoping about.
The fast-feedback machinery makes this concrete rather than abstract. Glucocorticoid feedback at the PVN works by triggering endocannabinoid release that suppresses presynaptic glutamate onto CRH neurons — so the route by which cortisol tells the controller to stand down is the same route that is already damaged. Supplying cortisol does not merely fail to correct reduced excitatory gain at these synapses; the mechanism through which it acts is reduced excitatory gain at these synapses. Whether transient non-genomic suppression accumulates into anything structural is unknown, and that is the load-bearing uncertainty rather than a detail.
The reading that would save it is that the lesion is on the input side — reduced excitatory gain onto PVN neurons, rebuilt by use — while the feedback acts on output, so load could strengthen the afferents while cortisol suppresses the efferent signal and the two never collide. I think that is plausible. I also notice that it is the reading I would prefer, which is a reason to distrust it.
Three things follow that I had not seen before writing this down.
My delivery method optimises against my own goal. I smoothed the dosing deliberately, because a slow structural process wants sustained input. But a feedback controller registers deviation, not level, so a flat and adequate curve is exactly what removes the error signal. Those two design aims are in direct tension and my current regimen resolves it entirely in favour of one. The variant worth trying is deliberate troughs — adequate on average, with predictable windows where cortisol is allowed to fall and demand can exceed supply while plasticity substrate is still present.
The two drugs are not symmetric. DHEA does not suppress ACTH. So it supplies substrate at no cost in feedback, where hydrocortisone supplies substrate and the suppression together. If the goal is to keep the controller honest, that argues for leaning on the arm that does not lie to it.
But cortisol is what makes the load possible, which is the knot. The clean version of the advice above would be to lean on DHEA and withdraw the hydrocortisone, since one raises drive and the other substitutes for it. The dissociation described earlier says that will not work as stated. Without the cortisol I could think but I could not exercise — and exercise is the load, and load is the demand signal that is supposed to drive the whole recovery. So hydrocortisone simultaneously enables the demand signal and suppresses the error signal. Remove it and you keep the error signal but lose the capacity to generate anything for the controller to respond to.
Which suggests the answer is timing rather than ratio. Cortisol present during the effort, because that is what makes the effort possible, and allowed to fall afterwards, so the deficit registers while the demand is recent. That is a sharper version of the deliberate-troughs idea: the trough should be placed relative to the load rather than scattered through the day. I have not tried it and it may be wrong. It is the first arrangement I can construct that does not defeat itself.
A taper should probably be done under load, not under rest. The standard advice when coming off a glucocorticoid is to take it easy. If the error signal is the mechanism rather than the side effect, that is backwards: withdrawing while demand is high gives the controller the largest possible gap between what is being asked and what is arriving, at a moment when the substrate for remodelling has not yet gone. Withdrawing while resting gives it the smallest. I have not done this and I am not recommending it. It is what my own argument implies, which is a different thing.
Three things, which I did not know when I started and which I would want anyone reading this to know before treating it as a model. A fourth point, on how the risks should be weighed at all, follows them.
Dosing late in the day is the most axis-suppressive schedule available, and I do not know whether that matters at this dose. Glucocorticoid taken in the evening blocks the overnight rise in ACTH that produces the morning cortisol peak, and does so far more effectively than the same amount taken earlier: split a fixed daily dose toward the evening rather than the morning and you get markedly more suppression from identical total exposure. Standard replacement practice puts the last dose at least four hours before bed for exactly that reason. What I cannot tell you is whether the effect is meaningful at a fraction of a milligram, since the evidence for it comes from doses many times larger — the principle is well established, the magnitude down here is not. It is worth flagging anyway, because it is the one part of the schedule that could in principle work against the thing it is meant to fix: on this page's own argument, evening glucocorticoid is what holds a drive-side deficit down.
The suppression risk is smaller than the phrase "taking hydrocortisone" suggests, but it is not the interesting number. Guidance generally puts the threshold for meaningful adrenal suppression above a daily hydrocortisone equivalent of 15 to 25mg (European Society of Endocrinology & Endocrine Society, 2024). At 2.5mg a day this sits well under that, and spontaneous recovery of the axis is the norm at doses in this region. That is reassuring about the specific catastrophe people worry about, and it is not a claim that a quarter-replacement dose does nothing to a system this page argues is already under-driven.
Bloodwork does not check the thing you would want it to check, though there is a workaround. Hydrocortisone is cortisol — no assay separates the pill from your own — so a serum cortisol drawn while taking it largely measures the drug, and a result inside the reference range says nothing about whether your axis is intact or being suppressed. ACTH is the clean readout: neither supplement touches it.
DHEAS is confounded but not ruined. Supplementing raises it directly, so the absolute number is uninterpretable. But DHEA has essentially no glucocorticoid activity and so does not suppress ACTH, which means your own production carries on underneath the supplement rather than being switched off by it. On a stable dose, DHEAS settles at a plateau attributable to the supplement, and a sustained rise above that plateau is your own drive coming back. The trend is readable even though the level is not, and it is the cheapest recovery signal available.
The conditions are the whole thing, though. Hold the dose genuinely stable, take enough serial measurements to tell a trend from noise, and taper in response to a sustained rise rather than a single good reading — because if you taper while the number is still moving you have changed two variables at once and can no longer read either. A single draw tells you nothing in any case.
Everything above is written in the register medical caution defaults to, where the intervention carries risk and the alternative is neutral. The alternative is not neutral, and pretending otherwise does real damage.
The comparison that matters is not "small risk versus no risk." It is a small, quantified, largely reversible risk set against the continuing cost of being unable to work, exercise, or get through a day. That cost is not a baseline. It is an active harm and it compounds: deconditioning, lost income, a career that does not wait, a social world that narrows. If a quarter-replacement dose is what stands between someone functionally disabled and a life, the arithmetic is not close, and treating the intervention as the only entry in the ledger is a failure of accounting rather than an excess of caution.
This page's own model sharpens it. If the behavioural loop is real, waiting is not holding steady — low cortisol means less movement and less light, which means less drive, which means less cortisol. Doing nothing is a direction, and it points down. On that reading the conservative option carries a mechanistic cost that never appears on anyone's risk sheet.
None of that makes the risks fictional. What I am objecting to is a default that weighs one side of the scale carefully and the other not at all — and which, as the section above argues, is applied by people who have no framework for the condition being weighed.
Not deference and not boilerplate — these are the specific failure modes, stated because knowing them is what lets you decide anything for yourself.
My own status: n = 1, unblinded, and strongly motivated to believe it is working. The change has been in capacity rather than mood — I can train hard and carry a full day, which on the account above is not merely a nice outcome but the mechanism, since load is the input that would eventually hold the state up without help. Whether it can be withdrawn is the interesting question and I would like the answer to be yes. But I want to be clear about which part is the win. Being able to live a life today is the win. Coming off it later is a hypothesis.
Notes and arguments, written down because nobody else was going to. The author is a software engineer who spent years operating queueing systems with adaptive rate limiting, got a T/T result and a burnout, and found that the stress axis behaves like a badly tuned control loop. That is a source of useful intuitions and of one characteristic failure mode, which is mistaking a coherent model for a correct one. I have done it before on this subject and expect to again.
Corrections are welcome.