If you have read Horowitz and Taylor, you already know most taper apps hand-wave the pharmacology. This page does the opposite. Here is exactly how the schedule is built, where the numbers come from, what they cannot do, and what we honestly do not know.
The dose math is deterministic code. No AI ever touches a milligram.
This matters most, so it goes first. Your schedule is generated by a plain, auditable function: given your drug, your dose, your goal, and a pace, it returns the same steps every time. A language model never sets, nudges, or rounds a dose. The companion can talk with you about how a stretch feels, and it can offer to re-run the deterministic planner at a gentler pace, but the milligrams themselves come only from the equations below, and every result is re-validated against the safety rails before you ever see it.
Hyperbolic, shaped by receptor occupancy
Dose and receptor occupancy relate hyperbolically, not linearly, so equal milligram cuts feel easy at the top and slam you at the bottom. The planner reduces a fixed fraction of your current receptor occupancy each step, then converts that back to a dose on the drug’s own curve. The milligram cuts run larger where the receptors are near-saturated and shrink through the low-dose tail, so each step lands as a roughly even change in what your nervous system actually feels.
Even mg cuts → a cliff at the end Hyperbolic cuts → even ~10% steps
For SSRIs and SNRIs the curve is anchored to human PET occupancy data, the same fits Horowitz and Taylor built their tapers on. Each drug carries an ED50, the dose at half of maximal serotonin-transporter occupancy: escitalopram 1.5 mg, fluoxetine 2.7 mg, citalopram 3.4 mg, paroxetine 5 mg, venlafaxine 5.8 mg, duloxetine 7.9 mg, vortioxetine 8.5 mg, sertraline 9.1 mg, desvenlafaxine 14.4 mg. Below the last few milligrams, where occupancy falls off a cliff, the steps get their smallest.
Benzodiazepines and Z-drugs: one shared GABA-A curve, and an honest caveat
Benzodiazepines and Z-drugs all act at the same GABA-A benzodiazepine site, so they share a single occupancy curve expressed in diazepam-equivalent terms (ED50 near 25 mg diazepam-equivalent), and each drug’s curve sits on its own milligram axis through its Ashton equivalence. Here is the honesty the rest of the market skips: that curve is model-based. It is derived from primate flumazenil PET imaging converted into human diazepam-equivalents, not a measurement of your specific drug in you. It is a careful approximation of the science, which is why your own response always overrides it. When a drug lacks defensible human occupancy data, the planner falls back to a generic hyperbola on percent of current dose, the safe, bounded default.
Three paces, all inside the safe envelope
Every preset sits comfortably below the hard rails. The difference is how gentle, not whether it is safe.
Gentle
~7% of current receptor occupancy per step
about every 14 days
each milligram cut capped at 10% of the dose
Standard
~9% of current receptor occupancy per step
about every 12 days
each milligram cut capped at 11% of the dose
Steady
~11% of current receptor occupancy per step
about every 10 days
each milligram cut capped at 11% of the dose
Drugs without a measured occupancy curve step by a flat 6, 8, or 10% of the current dose instead, on the same intervals. Whichever preset you land on, the steps still shrink in milligrams as the dose falls.
The rails that cannot be crossed
These limits are enforced in the math itself, not printed as advice. No pace, no personalization, and no edit can produce a schedule that breaks them. They are grounded in the Ashton Manual, the Maudsley Deprescribing Guidelines, and the 5 to 10% every 2 to 4 weeks, 25% per 2 weeks consensus.
Never more than 12% in one step
No single reduction can remove more than about one eighth of your current dose, whatever the pace or anything layered on top.
Never more often than every 7 days
A reduction can never land less than a week after the last one. A same-day second cut is caught too.
Never more than 25% in any 14 days
Across any rolling two-week window the total reduction stays at or under a quarter of the dose.
Longer holds near the end
Once you drop below 20% of your starting dose, every step gets an extra 7 days of hold, because a small milligram change is doing a large occupancy change down there.
You step off from a low dose, not a cliff
For occupancy-curve drugs the final jump to zero happens near 10% receptor occupancy; for the rest, only once the next dose is within about 4% of where you started.
Every edit is re-checked
Any schedule that is regenerated or changed is run back through the same validator. If it cannot satisfy the rails, it is discarded, not shown.
It adapts in both directions
A written-once schedule cannot respond to a hard month. This one watches your post-cut symptom trend and moves, always through the same deterministic planner, always re-validated.
Auto-hold on a rough stretch
When your check-ins show symptoms rising after a reduction, the taper can pause itself. A hold only shifts dates later, so it is inherently safe, and you can turn it off in settings.
Soften, and firm
Struggling, with room to go gentler, earns a one-tap re-plan at a slower pace. A sustained good stretch (at least four recent check-ins, none in the high band, a low average, and not just after another change) can offer to pick the pace up. Notice the asymmetry: slowing down is easy to earn, speeding up is only ever an offer you choose. If a re-plan cannot satisfy the rails, it is thrown away rather than shown.
Risk forces caution, never speed
Pace is decided in one place, so a higher-risk person can never reach a faster plan through a side door. A history of seizures, pregnancy or breastfeeding, being over 65, heavy alcohol use, other sedatives on board, or repeated prior attempts all force the gentlest pace and flag the taper as one that needs a prescriber supervising it. Longer time on the drug pulls the pace gentler too: ten years or more forces the slowest, and three to ten years nudges it one step gentler than sensitivity alone would give. The math only ever moves toward caution.
What we will not tell you
The limits of this method, stated plainly
· This is not clinically proven to work, and we will never say it is. It is an engineering of an evidence-based method into software. The plan it produces is a draft to review and adjust with a prescriber, not a prescription.
· The encouraging outcome figures you see on this site (for example roughly 70% coming off in cohorts where most had failed before) come from the tapering-strip and hyperbolic-taper studies, not from Subside. They describe the method, not a promise about you.
· The GABA-A occupancy curve is model-based, converted from animal imaging. Individual pharmacokinetics, formulations, and sources vary. Your lived response always outranks any curve.
· Hold or slow down whenever you need to. Reducing too fast, or stopping abruptly, can be dangerous. Seek urgent care for any seizure, severe confusion, or thoughts of self-harm.
Educational information, not medical advice. Taper only with a qualified prescriber. In crisis, call or text 988 (US) or your local emergency number. Safety & crisis resources →