Overview
The feedback system is one loop with two instruments: the Caffeine Behavior Instrument (CBI) sets each Resident User's starting values at onboarding, and the Adaptive Feedback System (AFS) refines those values from reported outcomes for as long as the profile is in use. The CBI asks what the user expects to feel; the AFS asks what they did feel.
The loop runs in seven stages:
- CBI onboarding — up to 14 behavioral questions.
- Starting values derived — Caffeine Decay Rate, Peak-Zone Floor, Wake-Hours Ceiling, Bedtime Ceiling, plus confidence flags.
- Phase 1, Orientation — about one week of silent observation.
- Phase 2, Introduction — at the first brew of week 2: what we saw, a CBI validation check, a boundary review, and the enrollment offer.
- Phase 3, Calibration — short one-tap check-ins across four dimensions, on a three-layer cadence.
- Adjustment — answers are weighted and smoothed; boundaries move only in small, confirmed steps that the user can see and undo, and large disagreements become an informed-choice advisory.
- Ongoing operation — per-profile, opt-out safe, fed by every brew, plan and logged drink.
Three principles govern every stage:
- Reported outcomes, never observed. The device has no sensing. Everything it learns about how the user feels comes from the user.
- The system recommends, the user decides, the machine always brews when the user asks it to. A boundary that moves changes the advice, never blocks a brew. Every adjustment is visible and reversible, and when the model and the user disagree sharply, the user is told and chooses.
- No sensitivity tiers for Resident Users. Every Resident boundary, the Peak-Zone Floor included, is calculated from the user's own answers; the boundaries are the quantified expression of sensitivity. Tiers apply only to Guest profiles as a conservatism shortcut.
Stage 1 — The Caffeine Behavior Instrument
The CBI is a derivation instrument, not a preference form: up to 14 behavioral questions, ordered through the user's day, each feeding exactly one output. Sensitivity questions never touch the clearance calculation, and clearance questions never touch the sensitivity boundaries. Counts are tied to a drink of known caffeine content, and frequencies use named buckets rather than vague adverbs.
| # | Part of day | Question | Answer options | Feeds |
|---|---|---|---|---|
| 1 | Morning | What time do you usually have your first caffeinated drink? | Time picker | Load model start time |
| 2 | Morning | What do you usually reach for first thing? | Regular coffee (~95 mg) · Large/travel-mug coffee (~200 mg) · Espresso drink · Black tea (~45 mg) · Green tea (~30 mg) · Energy drink (~110 mg) · Something else | Caffeine per drink (C) |
| 2a | Morning, only if espresso | How many espresso shots in it? | One (~70 mg) · Two (~140 mg) · Three or more (~210 mg) · Not sure (assume two) | Refines C |
| 3 | Morning | How many drinks before you feel it kick in? | Just one · Usually two · Three or more · Out of habit, not for the lift | Peak-Zone Floor |
| 4 | Typical day | How many caffeinated drinks on a typical day? | One · Two · Three · Four · Five or more | Cup count (N) |
| 5 | Typical day | When do you have most of your caffeine? | Early-heavy · Even · Late-heavy | Spacing of the reference day |
| 6 | Typical day | Is your routine about the same day to day? | Pretty consistent · It varies a lot | Confidence flag |
| 7 | Afternoon | On a too-much day, how many drinks had you had by then? | Around two · Three · Four · Five or more · Never hit that point | Wake-Hours Ceiling |
| 8 | Afternoon | How often do you feel jittery in the afternoon on normal intake? | Most days · A few times a week · Once in a while · Rarely or never | Decay Rate nudge |
| 9 | Afternoon | If you have a coffee at lunch, can you still feel it by dinner? | No, it's worn off by then · A little · Yes, clearly · Not sure / I don't have coffee at lunch | Decay Rate nudge (how long a drink lasts) |
| 10 | Evening | What time do you usually go to bed? | Time picker | Bedtime projection endpoint |
| 11 | Evening | Latest caffeine that still lets you sleep well? | Time picker | Last-drink time (cutoff) |
| 12 | Evening | When you have caffeine later than usual, how's your sleep? | Fine · A little harder to fall asleep · It clearly keeps me up | Cross-check on Bedtime Ceiling |
| 13 | About you | Do you smoke cigarettes regularly? | No · Yes, daily · Occasionally · Vape/nicotine only | Decay Rate physical modifier |
First-drink time, daily pattern, afternoon jitters, the lunch-to-dinner question and smoking are there so the decay rate starts from a research-anchored value and the Bedtime Ceiling can be computed from the user's whole day, not just their last drink.
Why we don’t ask personal health questions. Other factors that affect how fast caffeine clears, such as pregnancy, hormonal contraception and some medications, are deliberately not asked. They are sensitive health questions that raise privacy concerns: storing them would make the platform hold personal medical information, and many users would rather skip a question than share it. Caffevolve is also not a medical tool. Smoking is the one exception because it has a large, well-documented effect and is easy to answer.
Why a separate duration question. The cup-count questions measure sensitivity: how much caffeine the user can hold. The lunch-to-dinner question asks how long one drink's effect lasts, which is closer to clearance. The cutoff answer is not used for the decay rate. It mixes clearance with sensitivity and habit, and it already builds the Bedtime Ceiling, so using it again would count it twice.
When the answers need more context. If an answer gives no usable number ("out of habit" or "never hit that point"), or the two cup-count answers would leave a Comfort Zone narrower than 40 mg, the limits fall back to default values and are flagged for early feedback, so check-ins replace them quickly.
Science as a guardrail. If a computed boundary lands more than 50% outside the range the research supports, the CBI shows an informed-choice advisory: it notes the variance and asks the user to confirm their answers. If they stand by them, the platform honors the boundary. If not, it offers guided adjustments.
Stage 2 — Deriving the starting values
The CBI produces four numbers in a fixed order: the Caffeine Decay Rate first, then the two sensitivity boundaries, then the Bedtime Ceiling, which depends on the rate. The caffeine decay rate is set first, as a research-based starting estimate adjusted by three onboarding answers (smoking, afternoon jitters and the lunch-to-dinner question). The Bedtime Ceiling is then calculated from that decay rate and six onboarding answers describing the user’s usual day (first-drink time, what they drink, drinks per day, daily pattern, bedtime and cutoff). It is never the reverse, because only the half-life has a research anchor.
Caffeine Decay Rate (half-life)
Caffeine's half-life in healthy adults averages about 5 hours, with individuals ranging from roughly 1.5 to 9.5 hours. The platform starts at 6.0 hours, deliberately above the mean so the estimate errs toward protecting sleep. It applies smoking, then jitters, then the lunch-to-dinner answer, then holds the result within 2.0–10.0 hours. With these adjustments the reachable starting range is 2.75–7.5 hours.
| Input | Answer | Adjustment |
|---|---|---|
| Smoking | No / vape or nicotine only | 0 |
| Smoking | Occasionally | −1.0 h |
| Smoking | Yes, daily | −2.0 h |
| Afternoon jitters | Rarely or never | −0.75 h |
| Afternoon jitters | Once in a while | 0 |
| Afternoon jitters | A few times a week | +0.5 h |
| Afternoon jitters | Most days | +1.0 h |
| Lunch coffee by dinner | No, it's worn off by then | −0.5 h |
| Lunch coffee by dinner | A little | 0 |
| Lunch coffee by dinner | Yes, clearly | +0.5 h |
| Lunch coffee by dinner | Not sure / no coffee at lunch | 0 |
The lunch-to-dinner adjustment is deliberately small, about the weight of the jitters answer, because how long a drink is felt still reflects some sensitivity. Examples: a non-smoker with occasional jitters who feels a little of a lunch coffee by dinner stays at 6.0 h; a daily smoker lands at 4.0 h (2.75 h if also rarely jittery and a lunch coffee has worn off by dinner); a non-smoker jittery most days who clearly still feels a lunch coffee by dinner reaches 7.5 h. The adjusted rate, never the 6-hour base, feeds everything below.
Peak-Zone Floor and Wake-Hours Ceiling
Both are calculated directly from the cup-count answers against C, the caffeine in the user's usual drink.
The band between them is the user's Comfort Zone. A user who says two regular coffees (95 mg) to feel alert gets a Floor of about 143 mg, just short of the drink that does it for them.
Bedtime Ceiling
The CBI replays the user's usual day: N drinks of C mg from first-drink time to the cutoff, spaced by the pattern answer, each decaying at the adjusted rate. Whatever is still active at usual bedtime is an amount the user already sleeps well with, so that residual is the ceiling.
Worked example: 3 drinks × 95 mg at 7:00 AM, 11:00 AM and 3:00 PM, bed at 10:30 PM, T = 6 h. Residual at the 3:00 PM cutoff is 192.6 mg; projected 7.5 hours to bedtime it is about 81 mg. Because the ceiling is projected at the adjusted rate, the lunch-to-dinner answer moves it too: about 73 mg for "worn off by dinner" (5.5 h) and 89 mg for "yes, clearly" (6.5 h). That is the starting Bedtime Ceiling, held within a conservative range set at launch and tuned from sleep outcomes. It is a set figure that changes only gradually, through user feedback. At each brew, the platform works back from it to determine how much caffeine the new cup can carry.
Confidence flags
Two answers set no number. They tell the AFS where to look first.
- Consistency. "It varies a lot" widens the uncertainty on C and the whole load model, so feedback carries more of the weight for that user.
- Late-caffeine sleep. "It clearly keeps me up" alongside a high derived ceiling is a contradiction, flagged for early attention. It does not move the ceiling directly, because the cutoff answer already carries sleep sensitivity and nudging again would double-count it.
Contradictory or edge-of-range answers (very few drinks to feel alert and very few to feel wired) put the profile into a higher-sensitivity early-calibration mode, with low-confidence boundaries marked as first refinement targets.
Optional daily caffeine limit
Separately from the CBI, a Resident User may set a daily caffeine limit in milligrams on the Comfort Zone screen (SU9), or later on the Profile Summary. It is offered for users who have their own reason to cap caffeine, such as pregnancy, without the platform having to ask about it. It is optional, worded neutrally, and the platform never asks why it was set. Leaving it blank turns it off. Guests have no limit.
- What it counts. Total caffeine consumed, in raw milligrams with no decay, from bedtime to bedtime: the day starts 24 hours before tonight's declared bedtime. Device brews and drinks logged in CaffevolveGo all count.
- How it is used. It is a third test alongside the two ceilings. Each one sets a caffeine limit on the new cup: the caffeine still allowed under the Wake Hours Ceiling, under the Bedtime Ceiling, and under the daily limit. The lowest of the three decides the blend. Plan my Day and Plan Check treat "total for the day within the limit" as one more condition.
- Advisory, like the ceilings. The user is told when a drink would go over a ceiling or the daily limit, and can still brew or order it. The machine always brews when the user asks it to.
- Not a feedback boundary. The limit is the user's own choice, not derived from answers, so the AFS does not ask about it or adjust it. Only the user changes it.
Guests are different by design. Guest profiles skip the CBI and use fixed presets for every value, including the floor (40 / 40 / 60 mg for Higher / Moderate / Lower sensitivity).
Stage 3 — Phase 1: Orientation (about week 1)
For roughly the first week the AFS asks nothing. It observes silently, and the user is not told Orientation is under way. From the shared daily log it collects brew count per wake cycle, serving sizes, brew timing, first and last brew times, and any comfort-zone boundary proximity events that happen in normal use.
This becomes the behavioral baseline that early feedback is measured against. Every brew adds dose and timing data that decay-rate refinement will use once development is complete (in development as a top priority for the next platform version; see Stage 6). No boundary moves during Orientation.
Stage 4 — Phase 2: Introduction (first brew of week 2)
At the first brew of week 2 the AFS runs a four-screen sequence. It is a validation moment, not just an enrollment offer: the user sees what the platform learned, checks it against their CBI answers, corrects any ceiling, then decides whether to join check-ins.
| Screen | Title | What happens |
|---|---|---|
| P2-OBS | Here's What We've Learned | Week 1 baseline in plain words: typical daily brews, usual serving size, first and last brew times, any near-ceiling brews ("we noted it", never alarming), and the current decay-rate estimate, "set from your setup answers." |
| P2-VAL | One Week In — A Quick Check | CBI validation questions, one at a time in the CBI's own format (see below). An answer suggesting a miscalibration marks that boundary for early refinement. |
| P2-BDY | Boundary review | Shows the Peak-Zone Floor, Wake-Hours Ceiling, Bedtime Ceiling and Caffeine Decay Rate. The user may adjust the two ceilings; changes apply at once, and Calibration starts from these user-validated values. |
| P2-ENR | Enrollment offer | What check-ins involve (short, one-tap, at the right moment), then "Count me in" or "Not right now." |
The P2-VAL questions map one-to-one to the CBI boundaries:
- Floor: Over the past week, did the number of drinks it took before you felt alert and focused match what you told us at setup?
- Wake-Hours Ceiling: Did your overstimulation point, where you felt wired or jittery, feel consistent with what you told us?
- Bedtime Ceiling: Was your cutoff time accurate, or did you find yourself brewing later without it affecting your sleep?
The remaining inputs need no question. Smoking is a fact rather than an experience, and the user's daily pattern is checked automatically: the week-1 log shows when they actually drink, and the platform compares it with their onboarding answer.
The opt-out arc
Participation is always voluntary, and the device runs normally on stored boundaries whether or not the user enrolls.
| Opt-out | Wait | What happens next |
|---|---|---|
| 1st | 30 days | Warm re-ask screen, "Your best brew is still ahead of you," with sleep, focus and jitter benefit cards |
| 2nd | 90 days | Final re-ask, softer tone |
| 3rd | Permanent | No further prompts; re-enrollment only through Settings, with a condensed introduction |
Opting out pauses prompting only. All accumulated signal history is kept, so a returning user resumes refinement where they left off.
Stage 5 — Phase 3: Calibration check-ins
Once enrolled, the user gets short one-tap questions across four dimensions. Each question is timed to the moment its answer is freshest, and each refines one boundary. Refinement starts from the Phase 2 user-validated values, not the raw CBI output.
Where and how questions are asked
Questions appear at the user's next natural touchpoint: on the device at brew time, or in CaffevolveGo when the user is away. A push notification is used only as a backup when neither happens within the question's window. Every answer is a single tap on a three-icon set matched to its question, which needs no translation across markets.
The four dimensions
| Dimension | Question | When it's asked | Refines |
|---|---|---|---|
| D1 Sleep | How did you sleep last night? | First brew of the morning | Bedtime Ceiling (and, once development is complete, the decay rate) |
| D2 Focus | Did this brew get you going? | Mid-morning, after the first brew has taken effect | Peak-Zone Floor |
| D3 Jitteriness | Feeling wired or jittery? | Mid to late afternoon | Wake-Hours Ceiling |
| D4 Sleep onset | Did caffeine affect how easily you fell asleep? | Evening brew or last brew of the day | Bedtime Ceiling (and, once development is complete, the decay rate) |
D1 and D4 triangulate the Bedtime Ceiling from two angles: overall sleep quality, asked at the first brew of the morning, and how easily the user fell asleep, asked at the evening brew or last brew of the day. When they agree, the combined signal counts more (Stage 6).
The three-layer cadence
Three triggers decide when a question is due. Time-of-day anchoring then decides when it is asked: a due question waits for its dimension's window, so nobody is asked about last night's sleep at 3 PM.
- Boundary-triggered. Fires when a brew lands near a Comfort Zone boundary, the moment an answer says most. A brew that grazes the Wake-Hours Ceiling triggers D3; a brew whose projected bedtime residual approaches the Bedtime Ceiling triggers D4 at the evening brew or last brew of the day and D1 the next morning.
- Brew-count. A rolling floor cadence per dimension, regardless of boundaries, so a well-calibrated user who never pushes a ceiling still gets checked. Once development of decay-rate refinement is complete, it will be the main driver of it in early use, with D1 and D4 on an accelerated cadence. The interval resets whenever a boundary trigger fires for the same dimension.
- FR-Meta, every 60 days. A calendar prompt asking whether check-ins feel right: Too often, Just right, Not sure. It surfaces fatigue before it becomes an opt-out.
Brew-count intervals by profile stage:
| Dimension | Early Calibration (first 30 brews) | Active Calibration (brews 31–100) | Mature Profile (100+ brews) |
|---|---|---|---|
| D1 Sleep | every 3 brews | every 7 brews | every 14 brews |
| D2 Focus | every 5 brews | every 10 brews | every 20 brews |
| D3 Jitteriness | every 5 brews | every 10 brews | every 20 brews |
| D4 Sleep onset | every 4 brews | every 8 brews | every 16 brews |
During Early Calibration, D1 is also asked at the first brew of every morning, regardless of boundary proximity, so decay-rate data accumulates fastest when the model is least accurate.
Respecting the user's attention
The brew-count layer is governed by engagement. Every two weeks the platform looks at the share of check-ins the user answered and adjusts the brew-count pace by no more than two prompts a week:
| Answered in the last 2 weeks | Brew-count pace |
|---|---|
| 80% or more | may increase |
| 50–79% | holds |
| Below 50% | reduces |
| Below 25% for two periods running | FR-Meta is asked early |
Boundary-triggered questions are not throttled, because they are the most valuable and the least frequent.
Stage 6 — Turning answers into adjustments
No single answer moves a boundary. Each answer is a signal, weighted by how informative it was and smoothed into a running score per dimension. A boundary moves only when that score crosses a threshold, and then only by a small step. The values below are the platform's starting parameters, tuned against test profiles during validation.
Weighting and smoothing
Each answer is scored (for example, "slept poorly" = −1) and multiplied by a confidence weight from 0.4 to 1.0. The weight is highest after a brew that came close to the boundary, lowest for a routine brew-count prompt. The weighted answer then feeds an exponential moving average (EMA), a rolling memory in which recent answers count more than old ones.
| Profile stage | Brews | α | Behavior |
|---|---|---|---|
| Orientation | week 1 | not applied | observation only, no updates |
| Early Calibration | first 30 | 0.35 | fast adaptation |
| Active Calibration | 31–100 | 0.20 | moderate |
| Mature Profile | 100+ | 0.10 | slow drift correction |
Update rules
- Minimum evidence: at least 3 consistent answers in the same direction before a boundary moves.
- Maximum step: no single update moves a boundary more than 10% of its current value.
- Hard floors: Bedtime Ceiling never below 20 mg, Wake-Hours Ceiling never below 100 mg, Peak-Zone Floor never below 15 mg.
- Governor band: however much evidence accumulates, a boundary stays within a set band around its onboarding value.
- Agreement bonus: D1 and D4 answers that point the same way get a 1.4× weight.
- Ambiguity penalty: a likely-confounded answer, such as poor sleep with no high-residual brew behind it, gets 0.5× weight and needs more confirmation.
- Correlated answers: poor sleep and high jitters on the same day touch both ceilings and are weighted together, not routed separately.
- Low-confidence boundaries first: boundaries flagged at the CBI or in P2-VAL are prompted and refined before the others.
- Stepwise: after a move, the effect is checked at the next boundary-grazing brew before any further step.
- After a long break: a profile that returns after an extended gap keeps its refined values and resumes at Early Calibration pace, so it relearns quickly if habits have changed.
Every change is visible
When a boundary moves, the user sees a short note the next time they brew, for example: "We lowered your Bedtime Ceiling slightly based on your last few nights. Undo?" The change is already in effect; one tap reverses it. The user can also change either ceiling manually at any time. AFS refinements inform the user's judgment and never override it.
Decay-rate refinement comes first
In development, and a top priority in the next platform version. In the current release the decay rate stays at its CBI starting value.
The decay rate underlies every projection, so a wrong rate biases every recommendation; a wrong boundary errs in one place only. Once development is complete, refinement will start silently in Orientation and stay the top priority through Early Calibration. With real brews and logged drinks the platform knows the actual dose and time of everything consumed. Each night it predicts the bedtime residual under the current half-life, and the D1 and D4 sleep answers label it. A handful of nights such as "predicted 45 mg, slept fine" versus "predicted 45 mg, kept me up" is enough to fit the half-life that best separates good nights from bad. As the rate becomes more accurate, every blend recommendation becomes more accurate with it. The Bedtime Ceiling still changes only through sleep feedback.
When the model and the user disagree: the 50% advisory
If reported outcomes diverge from the model's predictions by more than 50% for a dimension, averaged over a rolling 14 days, the AFS does not adjust on its own. It shows an informed-choice advisory: what the model predicted, what the user reported, how large the gap is, and an invitation to consider a boundary change. The user can accept, decline, or ask for more time. While the advisory is active, no boundary moves without the user's acknowledgment. The science advises; the user decides.
Worked example — D1 Sleep
- Trigger. Early Calibration, Bedtime Ceiling 55 mg. An 8:30 PM brew projects a 52 mg bedtime residual, 5.5% from the ceiling, so it qualifies as boundary-triggered.
- Question. At the 7:15 AM first brew, D1 asks how they slept. Answer: "Slept poorly" = −1.
- Weight. Proximity 5.5% gives a confidence weight of 0.92, so the weighted answer is −0.92.
- Smooth. EMA updates at α = 0.35. It is answer 1 of the 3 required, so the signal is logged and nothing moves.
- Move. Two more consistent negative D1 answers that week cross the threshold. The ceiling drops 8%, inside the 10% cap: 55 mg to 51 mg. At the next brew the user sees the note and can undo it.
Stage 7 — Ongoing operation
After calibration the loop keeps running at Mature Profile pace, so boundaries follow the user as habits, tolerance and circumstances change.
- Per profile, not per device. Enrollment, phase and response history belong to each Resident User Profile. An Advancara household with five profiles has five independent AFS states, and a new profile never inherits another's refinements.
- The device never depends on feedback. Opted out or enrolled, it brews on the stored boundaries. Feedback improves it; it is not required.
- Complete intake makes answers meaningful. Every predicted residual a check-in is judged against includes device brews and drinks logged in CaffevolveGo through ECE. The ECE prompt that appears a couple of times a day before a brew keeps that record current.
- Plans count too. Feedback from days run on a Plan my Day plan or a Temporary Alternative Plan (TAP) is collected and used exactly as with single brews.
- The daily limit counts everything too. When a limit is set, the same complete record of brews and logged drinks is what it is checked against.
- Changes reach the next brew. An updated boundary feeds the next recommendation and the next Plan Check re-solve. Plan Check and the dual-ceiling check always use the current values.
- Guests are outside the loop. Guest profiles use fixed sensitivity presets and do not take the CBI.