Skip to content

ADR 0008: Goals, optional body data and "your daily guide"

Status: accepted · 2026-11-15 Supersedes: the public promise "no weight and no height" (docs/your-data.md, docs/guide/settings.md, PRODUCT.md "no biometrics") and the "a goal only moves the energy guide by 10 %" rule. Amends ADR 0006 and docs/no-dark-patterns-standard.md for the points below.

Context

The optional profile (age range, sex, activity, goal) gave a daily guide that could only be a rough reference: no height or weight, so no personal resting energy, no protein per kilogram, and a goal that nudged the figure by 10 %. The full application audit (reports/audit-2026-11-15.md) found that the goals did not do what a person choosing "lose a little weight" would expect, and that people who want to lose weight were left without help from a tool that holds their food data.

The maintainer decided on 2026-11-15 that the rule to protect is again pressure, not information. Body measurements the person chooses to enter are information they already know. What stays banned is a verdict on the person and a sense of an allowance being spent.

Decision

1. The personal energy figure is "your daily guide" (French: "votre repère du jour"). It is a reference to understand a day against, not a target to hit and not a limit.

  • Going past it is visible but neutral: same colour, plain text ("about 112 % of your daily guide").
  • One new, opt-in, per-person setting, off by default, in Settings → Nutrition: Show the distance to my daily guide. When on, one plain text line states the distance: "about 450 kcal to your daily guide", and once past it "about 120 kcal past your daily guide". No colour change, no icon, no countdown ring, no allowance-being-spent visual, and never in the coach's voice. It is stored as the key/value setting ui.showDistanceToGuide.

2. Optional body data is allowed. Height, weight (dated weigh-ins) and waist (dated), all optional, all inside the encrypted store, synced with the vault, included in the export and erased by Erase all local data and Clear profile. The profile gains heightCm, units (metric or imperial; storage is always metric), pace (gentle, standard, faster), customEnergyKcal, pregnantOrBreastfeeding and the activity level veryHigh. Weigh-ins and waist live in a new synced table, body_measurement (weight in kg, 20 to 400, stored to two decimals so that a weight typed in pounds comes back as the same pounds; waist in cm, 30 to 250, one decimal; one per calendar day per kind: a second entry the same day replaces the first). Plates gain a nullable meal slot (breakfast, lunch, dinner, snack, late). This ships as schema version 5. Older data stays readable: a snapshot tagged with a schema version below 5 (an .ancvault backup or a relay copy written by the previous app) is parsed with the v4 key set, with no measurements and every plate's slot null, and is a v5 snapshot once read; a v4-tagged snapshot that carries a v5 key is malformed, and a newer version is still refused.

  • Birth date stays banned. Age remains an age band.
  • BMI and the waist-to-height ratio may be shown as plain numbers next to the WHO and NICE reference ranges written as text ("Reference ranges for adults: 18.5 to 24.9 ..."), behind a disclosure, never as a label attached to the person ("you are overweight"), never colour-coded. The caveats are shown with them: a muscular build, age 65 and over, pregnancy, and the lower ranges of 23 and 27.5 for people of South Asian, Chinese, other Asian, Middle Eastern, Black African or African-Caribbean background.

3. The coach may help with weight loss when the person chose that goal. Implemented in lot 6. It never moves outside the person's chosen goal, never praises or blames a change in weight, and follows the safeguards below. What the coach may and may not say is listed under "What the coach may and may not say" below.

4. No server accounts (unchanged) and adaptive expenditure estimation is out of scope: the guide is computed from the profile, not from a person's logged intake and weight history.

Evidence defaults

The guide is an estimate, usually within about 10 %. It is worked out like this.

  • Resting energy. Henry (2005) equations from weight and height when both are known; the weight-only equations when only weight is known; a reference-weight fallback otherwise (EFSA 2013, scientific opinion on dietary reference values for energy). The age bands are 18-29 (Henry 18-29), 30-49 (Henry 30-59) and 65+ (Henry 60+); the 50-64 band straddles Henry's 30-59 and 60+ ranges (60-64 belongs to 60+), so it uses the mean of the two equations.
  • Activity (PAL). 1.4 (low), 1.6 (moderate), 1.8 (high), 2.0 (very high).
  • Goal adjustment.
Goal Adjustment
Lose A deficit by pace: gentle -250, standard -500, faster -750 kcal per day, capped at 25 % of maintenance (the cap is floored to the 50 kcal below, so it is never exceeded). With no pace chosen, gentle. Only when the BMI is known and at least 18.5.
Gain +250 (gentle or standard) or +400 (faster) kcal per day. Only when the BMI is known and below 30.
Maintain, eat better 0.
  • Floors. 1500 kcal (male or unspecified) and 1200 kcal (female). The guide is never set below the floor by a goal. When the deficit of the chosen pace would cross the floor but at least 250 kcal (the gentle step) still fits between maintenance and the floor, the guide is the floor and the explanation says so. When less than that fits, no deficit is applied at all: the guide stays at maintenance and the calm line under Safeguards is shown (the 250 kcal threshold is itself a convention). When the estimate of maintenance is itself under the floor, the guide is raised to the floor whatever the goal, and the screen says so with the floor wording ("the lowest figure used for a guide, a little above what keeps your weight steady"), never as an effect of "your goal and pace".
  • Protein. 0.83 g per kg of body weight; 1.0 g per kg at 65 and over; 1.2 g per kg when a deficit is applied. Computed on actual weight; on the weight at a BMI of 25 when the BMI is above 30 and the height is known; on the actual weight capped at the weight of a BMI of 30 at the reference stature of the sex when no height is known (1.76 m for men, 1.63 m for women, the reference statures behind the EFSA 2013 reference weights; the mean of the two weights for "prefer not to say"), so a large weight without a height cannot inflate it (300 kg gives about 77 g from the base rule, not 250 g); the reference weight when no weight is known. The figure is then bounded: at least 40 g, at most 160 g, and never more than 30 % of energy, so that carbohydrate stays at 35 % of energy or more.
  • Fat 35 % of energy; saturates 10 % of energy; carbohydrate the remainder; sugars 90 g scaled by energy over 2000; fibre 25 g (30 g is named as the aim in the explanation text only); salt 6 g, the EU label reference.

Sources named in the app and the docs: Henry (2005), the Henry equations as used by EFSA; EFSA (2013) energy opinion; EFSA (2012) protein opinion and ANSES; the AHA/ACC/TOS 2013 obesity guideline; NICE NG246 (2025) on overweight and obesity management; CDC guidance on adult BMI and healthy weight; and the NIH Body Weight Planner, a dynamic model in which weight loss slows over time.

Be plain about what these figures are. The floors of 1500 and 1200 kcal, and the rule-of-thumb "5 to 10 % of body weight in six months", are conventions from clinical guidance practice, not clinical thresholds. They are used because they are widely accepted starting points, and they are labelled that way.

What the coach may and may not say

With the goal lose (or gain), and while no safeguard holds the guide (pregnancy or breastfeeding, a BMI under 18.5, 65 and over), the coach, built-in or AI, may: explain energy balance plainly; compare the day or the week with the personal daily guide (never the EU reference intakes) from totals the coverage model trusts, at dinner or the end of the day's last meal, as information and a question ("Much of today's energy came from the pizza and the soda ... Would a smaller portion of one of them suit you next time?"); say that eating enough matters when a lose or gain day is well below the guide, never as praise; suggest satiety helpers (protein, vegetables, water in place of a sugary drink) naming foods the person actually logs; describe a realistic pace (0.25 to 0.75 kg a week at first, 5 to 10 % in six months, conventions as above) and that plateaus and day-to-day changes are normal; relate the weight trend to the logged energy descriptively; offer one small step.

It may not: set or suggest an intake below the guide or below the floor; recommend fasting, skipping meals, meal replacements, supplements, "detox", weight-loss drugs (or name a medicine, a drug or a supplement at all), or exercise as compensation; comment on the body or label the person; shame or moralise food ("bad", "cheat", "guilty"); diagnose; push weight loss under any other goal; praise a loss or blame a gain; use a streak, a score, points or a "remaining" figure. It points to a doctor or a dietitian for a medical condition, pregnancy, medication, being under 18, or a wish to lose more or faster. A request for a very low intake or a very fast loss (over a kilo a week) is answered on the device with a fixed calm line and is never sent. The thresholds are named constants (ENERGY_RULES, WEEK_RULES), conventions and not clinical thresholds: above the guide at 115 % (lose) or 130 % (other goals), below it at 70 % (lose, with three plates) or 80 % (gain), a weight trend described only over about four weeks. Food-quality remarks (pulses, fish, wholegrain, nuts, cured and red meat, sugary drinks) follow France's public-health guidance, are neutral and at most one a week.

The AI context gains, in lot 6, the week's average energy next to the rounded guide, the foods that carried the most energy, the accepted step and, only behind a separate, off-by-default switch (Share my weight trend with the coach) and the per-destination coaching consent, a coarse weight trend (direction and the change over about four weeks to the nearest half kilo). Weigh-ins, height, waist and BMI are still never sent; docs/ai-tiers.md lists the fields of each mode and tests pin them.

What enforces this, honestly. For the built-in coach the rules are in code and always apply. For an AI-written text there are three layers, none of them a guarantee:

  1. The system prompt asks the model for exactly this (and tells it that the guide it sees is rounded to 100 kcal, that it is the only daily figure it may state, and that logged amounts are described, never prescribed).
  2. The lexical backstop, validateCoachText (and detectSupportSignal for questions): a backstop of pattern and number checks that cannot catch every wording. On 2026-11-15 an independent measurement on 180 fresh sentences found it refused only 53 % of plausibly worded harmful outputs (it accepted 92 % of safe ones) and routed 33 % of unsafe questions to support (none of the ordinary ones). It was then extended with the wordings that measurement exposed and those sentences became regression fixtures; lexical rules are near their ceiling, so this is the cheap layer, not the strong one.
  3. A second-pass model check (ai-review.ts): the same AI, through the same rung, destination and coaching consent, classifies the question before the coach call (it can route it to the existing support copy) and the validated answer before it is shown, with a fixed rubric, temperature 0 and a strict JSON reply. It sees the text and, for an answer, four context values (language, whether weight talk is allowed, the rounded guide, the floor); for a chat turn it also sees the person's question and the previous question and answer, as separate delimited data-only blocks (random per-request nonce in the tags), so an answer is judged against what it answers. It fails closed: an error, timeout, rate limit, unreadable reply or an unsafe verdict means the text is not shown (a note keeps the built-in note; a paragraph or chat answer shows the calm "not sure" line). It has no off switch. It costs 2 to 3 calls per interaction instead of 1 (docs/ai-tiers.md, docs/self-host.md). It is only as reliable as the configured model, it can be wrong both ways, and it has not been measured against a real model.

A text that passes all three has only been found acceptable by three imperfect checks; the validator is never the thing that makes the built-in coach safe, and a text that fails any layer is replaced by the built-in coach's note.

The weight trend is sent only while weight help is allowed, that is a lose or gain goal that no safeguard holds, so a held guide (below the floor, no body data, no room for a surplus, as well as pregnancy, a low BMI and 65 and over) sends no trend even with the switch on, and the validator still compares figures with that held guide.

What stays banned

  • Birth date (age band only).
  • The banned vocabulary stays banned by default: maximum, limit, budget, exceeded, over, and remaining as pressure. The only place the wording "to your daily guide" and "past your daily guide" may appear is the opt-in distance line above.
  • Labels on the person ("overweight", "underweight", "obese", "healthy weight" as a verdict).
  • Colour-coding of BMI, of the waist-to-height ratio, of a weight or of any chart.
  • Countdown rings, progress rings and any allowance-being-spent visual.
  • The coach's voice for the distance line: the coach never quotes it.
  • Weekly weight-change targets used as scores, streaks of weigh-ins, weigh-in reminders by default, charts coloured good or bad.

Safeguards

  • The guide stays at maintenance, with no deficit, and a calm fixed line shows the pointer to support or a professional, when: the goal is lose and BMI is below 18.5; the guide could only be reached below the floor; the person is pregnant or breastfeeding; or the person is 65 or over with a BMI below 30. The line reads: "Your daily guide stays at what keeps your weight steady. If you want to lose weight, talking to a doctor or a dietitian first can help." For 65 and over: "After 65, weight loss is best decided with a doctor. Your daily guide stays at what keeps your weight steady."
  • A deficit needs a known body. With the goal to lose weight and no height or no weight (the BMI is unknown), the guide stays at maintenance and a calm notice asks for height and weight to work out a weight-loss guide. A BMI under 18.5 holds it too. The same for the goal to gain: the BMI must be known and below 30; unknown holds with a notice asking for the same, 30 or more gives no surplus.
  • The pregnancy and breastfeeding flag is ignored by the guide when the sex is male and is cleared when the person switches to male. Under the checkbox one calm line says the guide does not account for the extra needs of pregnancy or breastfeeding and that a midwife or a doctor is the right person to ask.
  • A custom energy figure set by the person is accepted from the floor up, with one exception: wherever a safeguard above holds the guide at maintenance (pregnant or breastfeeding, a BMI under 18.5, 65 and over with a BMI under 30 or unknown, the goal to lose weight with an unknown BMI), a figure below the maintenance estimate is refused in the form with a calm explanation and ignored by the guide (customRefused carries the reason), and the notice is shown; a figure at or above maintenance is accepted. The "How this is worked out" disclosure says what happened.
  • One weigh-in per calendar day per kind; a second the same day replaces the first.
  • Pace of the weight trend. When the weight trend (an exponentially weighted moving average on calendar days, 10 % per day, packages/core/src/value-types/weight-trend.ts) falls faster than 1 % of body weight per week over the last 14 days, one calm fixed line is shown: "Your weight is going down faster than about 1 % a week. A slower pace is easier to keep and kinder to the body; consider a gentler pace in Profile, or talk to a doctor or a dietitian." It is shown in three places, chosen on purpose: on Trends, next to the daily guide in Profile and on Today (Today is where the guide is read every day, and a safeguard is only worth having where the person looks). It links to the pace choice, is never colour-coded, is independent of the chosen goal, is shown with numbers hidden (it states no figure), can be dismissed for 14 days (ui.paceNoticeDismissedAtUtc, synced), and asks for nothing. The data it needs, all named in PACE_SAFEGUARD: the first weigh-in at least 28 days before today (the start of the average and the first weeks of a routine are mostly water: a 100 kg person who loses 2.5 kg of water in five days and then 0.5 kg a week is not shown the line, nor is a 60 kg person who loses 1.5 kg and then 0.3 kg), at least 3 weigh-ins spanning at least 14 days within the last 28 days (one a week is enough; it was four in 14 days, which a weekly weigher could never reach), and a latest weigh-in at most 14 days old, so the line does not outlive its data. The thresholds are named constants; like the floors, they are conventions, not clinical thresholds. The trend and its change over 7 and 28 days are described with "lower" or "higher than", never as progress, and there is no target weight. The change and the pace use the real span between the points compared: when no weigh-in lies within 3 days of the start of the window, the comparison is with the nearest earlier weigh-in and says so ("5.0 kg lower than 60 days ago"), and the pace is computed over that span.
  • Typing slips. The screen is named Trends (French: Évolution), never "Progress": a weight is not progress or its opposite. A weight typed that differs from the last known one by more than 10 % or 5 kg (whichever is larger) is asked about once, in the weigh-in sheet, in Profile and in the welcome: "That is quite different from your last weigh-in. Save it as typed?" with Save and Change it, repeating no figure. In the trend maths a weigh-in that differs by more than 15 % from the median of its neighbours (up to three before and three after, at least three of them) is left out of the trend, the pace, the safeguard and the chart axis (OUTLIER); it stays in the list and the table, marked "not used for the trend", with its Edit action. It is never deleted.
  • Earlier weeks. An earlier week's note, the dock's label and the coach entry say which week it is by its dates. Browsing an earlier week does not move where a new plate lands: today, unless the person picked a day in the strip. "This week" in the export is the current week.
  • No weekly weight-change targets as scores, no streaks of weigh-ins, no weigh-in reminders by default, no charts coloured good or bad.
  • Weight, height, waist and BMI are never sent to any AI. (A coarse weight trend, behind a separate off-by-default switch, is the one derived exception; see "What the coach may and may not say".) What the coach receives is "your daily guide, rounded, which is worked out from your weight and height if you entered them": the guide is encoded by one helper (coarseGuideJson) before it is sent, energy rounded to the nearest 100 kcal and protein and the other gram figures to the nearest 10 g, because the precise protein figure is a factor times the weight. The screens keep the precise figures. docs/ai-tiers.md and the in-app copy say exactly this. A later lot would need its own ADR to send more.
  • Discontinuities. The floor, the cap and the safeguards are thresholds, so the guide can move by a few hundred kcal for a small change of weight or height. No smoothing is applied; the disclosure says so honestly when the floor or a safeguard decided the figure.
  • The support line ("If food or weight worries take up a lot of your thoughts, talking to someone can help.", with the link) also sits next to the goal and pace choice, always (not only for a weight goal), and once in the Body section, covering weight, height and waist, besides the places listed in the standard.
  • Removing measurements. Clear profile clears the answers only. Remove all measurements is a separate action behind the app's confirm dialog; it names what is removed and that it applies on every device that syncs, and it deletes through the store so that rows another device just synced go too.
  • Numbers off. With nutrition numbers turned off, the Profile's guide section shows no kcal, protein, maintenance, BMI or waist-to-height figure; the notices stay, with a line saying figures are hidden and a link to Settings, Nutrition.

Consequences

  • The public promise changes. "No weight and no height" is removed from docs/your-data.md, docs/guide/settings.md, PRODUCT.md (Users, Positioning, Capabilities, Principles), docs/features.md, docs/faq.md, the README and the website text. What stays true: no name, no birth date, no account, nothing readable by the server.
  • More health data. Weight together with what a person eats is health data under GDPR Article 9. docs/self-host.md carries a short non-legal note for operators. The server only ever sees ciphertext.
  • Adults only is unchanged. Age is still an age band; body measurements are not age and not an identity. "No biometric identification" is unchanged: height, weight and waist are measurements the person types, not biometric identifiers.
  • Sync, export and erase cover the new data. The new synced table is body_measurement; the new key/value keys include ui.showDistanceToGuide. Merge is by uid like every other table (ADR 0007), the JSON export includes them, and erase removes them.
  • The AI context gains nothing about the person's body by default: the daily guide it already carried is now sent rounded (see Safeguards), and the copy says it is worked out from the weight and height if entered. Lot 6 adds, behind its own switch only, a coarse weight trend (see "What the coach may and may not say").
  • Tests. Copy tests pin the opt-in distance wording to its one place and keep the banned vocabulary banned everywhere else; calculation tests pin the equations, floors, caps and safeguards; day-boundary and daylight-saving tests cover "one weigh-in per calendar day" and the meal slot.
  • Risk. Weight can be a sensitive subject. Mitigation: everything is optional, the safeguards above, the support line beside every body-data entry point, and the named-reviewer gate on every copy change.