Minimente Intelligent Breathing System

Document type: Addendum to PRD v4.0 — standalone, mergeable Applies to: Minimente — Stress Management (Programme #1) and all future programmes Addendum version: A1.0 — DRAFT FOR REVIEW Date: 2026-07-27 Owner / approver: Minni (founder, content author) — all content and clinical-framing decisions Intended audience: Implementing engineering team (incl. coding agents), design, and the founder for content sign-off Status of PRD v4.0: Frozen. This addendum does not modify it. Where it changes the effect of a v4 requirement, that is stated explicitly in §A0.2 and recorded as a decision for Minni, not as a fait accompli.

---

A0. What this document is, and what it touches

A0.1 Why it exists

PRD v4.0 contains one breathing exercise: FR-M1-04, "Breathing Reset (4-4-4-4 Box Breathing)", taught in Module 1, mirrored into the Emergency Tools as "Breathing reset" (FR-ET-03/FR-ET-08), and available for Toolkit accrual (FR-CORE-07). The product owner's position — which this document tested against the literature and largely, though not entirely, confirms — is that a single breathing exercise is the wrong product design, and that Minimente's core principle (the most appropriate support for the current situation) should apply to breathing as it applies to everything else.

This addendum specifies an Intelligent Breathing System: a small, evidence-graded protocol library; a deterministic, explainable recommender expressed as content data; a player that works silently, without audio, and without motion; a safety model in which contraindications shape the defaults rather than decorating an appendix; and integration as a reusable subsystem across Toolkit, "Right now", Search, Exercise Library, Programmes and Emergency Tools.

It is written to be merged by an implementing engineer with minimal editing. It contains no application code.

A0.2 v4 sections and requirements this addendum touches

Nothing in this addendum closes any v4 open question. It opens eight new ones (§A10).

A0.3 Research provenance and confidence

Five parallel research strands were commissioned — scientific literature, clinical practice, performance & human factors, product landscape, UX & personalisation — and all five completed. The literature and performance strands returned late, after a first draft of this document had been written against a shorter targeted search; their findings materially corrected that draft and are reflected in the current text. The three corrections that changed conclusions are called out where they occur, because they are the kind of thing a reader deserves to see rather than have smoothed over:

1. A 400-participant placebo-controlled RCT found no effect of coherent/resonance breathing over a credible breathing placebo (§A2.5). This demoted BR-3's rationale from "rate-specific physiology" to "longer dose of the same thing", and it is the single most important finding in this document.
2. Box breathing lost a clean physiological head-to-head against 6 breaths/min, and lost again on cognitive-task performance against extended exhalation (§A2.2). This strengthened, rather than created, the demotion at §A3.2/BR-5.
3. The exhale-emphasis advantage is contested, not settled — the tightest test of it is null (§A2.1). The library still leads with a longer exhale, but on preference and safety grounds as much as evidential ones.

Residual limitations, stated rather than buried: several primary sources were paywalled or 403-blocked and were verified at abstract/record level only (flagged inline as [abstract-level]); pooled effect sizes for one meta-analysis could not be extracted; and no completed meta-analysis of breathing and cognition in healthy adults exists at all. OPEN-BR-08 retains a single confirmatory clinical pass before build, folded into the OPEN-27 review that must happen anyway.

---

A1. Product philosophy for breathing

The system answers "What breathing exercise is most likely to help this person right now?" — not "Which breathing exercise would you like?". Five principles bind it.

• A1-P1 — Evidence sets the library; need sets the presentation. The protocols that ship are chosen by evidence quality and safety. What changes between situations is mostly dose, duration and framing, not protocol identity (§A2.5 — this is the central and least comfortable finding of the research).
• A1-P2 — Deterministic and explainable, always. Every recommendation is produced by an ordered, declarative rule list and carries a human-readable reason. No ML, no embeddings, no conversational AI (§4.3, §9.4 remain binding).
• A1-P3 — Non-clinical, non-diagnostic vocabulary only. The system reasons over clinical evidence internally; it never surfaces a clinical label. §7.7's house vocabulary is extended, not replaced (§A4.1).
• A1-P4 — Safety shapes defaults, not footnotes. A protocol's contraindication profile determines whether it can be a default at all (§A5).
• A1-P5 — Restraint is the feature. A large library is not free (§A9) and, per §A2.5, is not evidenced. The recommended v1 library is five protocols, one of which the product already owns.

---

A2. Evidence review

Graded strong (multiple systematic reviews/meta-analyses converging, or consistent large trials), moderate (one good meta-analysis, or ≥1 well-conducted RCT with converging mechanism data), weak (small/uncontrolled/single studies, or contested), insufficient (no protocol-specific evidence located).

A2.1 What is well supported

Slow breathing → acute autonomic change — strong. This is the only genuinely strong claim in the area. - Laborde et al. (2022), Neuroscience & Biobehavioral Reviews 138:104711 — systematic review and meta-analysis, 223 studies; vagally-mediated HRV increased during breathing, immediately after a single session, and after multi-session programmes. Authors describe it as low-tech, low-cost, with minimal adverse effects. (Pooled numeric effect sizes are in the paywalled full text and were not extracted.) - Shao, Man & Lee (2024), Mindfulness — meta-analysis, 31 studies, n = 1,133: systolic BP SMD = −0.45; RMSSD SMD = 0.37; SDNN SMD = 0.77. But negative-emotion reduction was only marginal: SMD = −0.51, p = 0.06 — non-significant. The authors' own summary is "reliable" for cardiovascular function and "modest" for emotion. - Zaccaro et al. (2018), Frontiers in Human Neuroscience 12:353 — 15 studies from 2,461 screened. Caveat rarely quoted: it contained no RCTs (10 within-subject, 5 pre-post), computed no effect sizes, and could not meta-analyse. It is a mechanism review, not an efficacy review.

Slow breathing → self-reported stress/anxiety — moderate, with small effects that shrink under proper control. - Fincham, Strauss, Montero-Marin & Cavanagh (2023), Scientific Reports 13:432 — meta-analysis of RCTs; stress g = −0.35 [−0.55, −0.14], 12 RCTs, n = 785; anxiety g = −0.32; depression g = −0.40. Zero low-risk-of-bias studies (3 high, 9 "some concerns"); all outcomes self-reported. Slow-paced subgroup significant; fast-paced (2 studies) not. The authors warn explicitly against "a miscalibration between hype and evidence" — a phrase worth keeping in front of whoever writes the marketing. - Leyro et al. (2021), Clinical Psychology Review 84:101980 — 40 RCTs, N = 3,047. Overall g = −0.550; against inactive controls g = −0.709, but against active controls g = −0.279. Panic subgroup g = −0.784 non-significant (k = 6). Neither duration nor session count predicted outcome. The active-control figure is the honest one for a product that will be compared to "taking five minutes off". - Goessl, Curtiss & Hofmann (2017), Psychological Medicine 47:2578–2586 — HRV-biofeedback meta-analysis, 24 studies, n = 484, between-group g = 0.83. Two caveats that matter: it is device-mediated, and the app-relevant version is weaker — Vann-Adibe et al. (2025), Applied Psychophysiology and Biofeedback, 18 RCTs, n = 1,352 remote HRVB: depression g = −0.41, HRV g = 0.44, stress non-significant (p = 0.152). Wells et al. (2012), PLoS ONE 7(10):e46597 (n = 46 musicians) found slow breathing with and without biofeedback equivalent — so Minimente can claim the slow-breathing mechanism without sensors, and must not borrow the lab-biofeedback effect size.

Workplace stress recovery — moderate, and the most directly transferable evidence the product has. - Chelidoni et al. (2020), JMIR mHealth and uHealth 8(10):e19412 — n = 75 full-time employees, lab stressor, 5 minutes of app-guided breathing at 6 bpm during recovery vs mindfulness body scan vs control. Breathing produced significantly better HRV recovery (RMSSD p = .002) — beating a mindfulness comparator. - Plans et al. (2019), JMIR Formative Research 3(1):e12227 — n = 75, 6-minute guided diaphragmatic breathing post-stress; better HRV recovery, medium-to-large effects. But subjective stress returned to baseline in all three groups including control. Physiological recovery ≠ felt benefit, and the product must not promise the latter.

Exhale emphasis — contested. Weaker than the popular account, and weaker than this document's first draft assumed. - For: Van Diest et al. (2014), Applied Psychophysiology and Biofeedback 39(3-4):171–180, n = 30 — a lower inhale:exhale ratio produced more relaxation, less stress and higher HF-HRV, though only at the slow rate. Balban et al. (2023) — cyclic sighing had the best composite profile of the four arms. - Against: Meehan & Shaffer (2024), Applied Psychophysiology and Biofeedback 49(3):407–417 — two within-subject RCTs (n = 26; n = 16 replication), 1:1 vs 1:2 at 6 bpm: null on RMSSD, SDNN, pNN50, LF power and sample entropy. Their review of nine prior studies splits 4 favouring longer exhale / 3 no effect / 1 favouring equal / 1 favouring longer inhalation. They recommend 1:1 by client preference. - Honest position: the longest-standing rationale for the product's default protocol is plausible, popular, and empirically unsettled, with the tightest test null. BR-1 still leads the library — but on safety (no holds), simplicity, and public-health precedent as much as on evidence of superiority.

Attention/cognition — weak, and dose-dependent in a way that rules out short sessions. - Laborde et al. (2022), Journal of Psychophysiology 36(1):13–27 — n = 78, 3 × 5 min slow-paced breathing improved Stroop interference accuracy and operation span, but the effect was not mediated by RMSSD — the assumed vagal mechanism was not demonstrated. - Blaser, Weymar & Wendt (2023), Frontiers in Psychology 14:1292983 — n = 56; a 5-minute HRV-biofeedback session raised vagal HRV by ~1.5 SD but produced a clean null on attentional control, explaining 0.5% of unique variance. The authors note every prior study finding acute cognitive benefit used 10–17 minutes. - Ma et al. (2017), Frontiers in Psychology 8:874 — n = 40, 20 sessions over 8 weeks at ~4 bpm with device feedback; improved sustained attention, lower cortisol. Routinely miscited as evidence for "diaphragmatic breathing"; it is a slow-breathing-plus-biofeedback training study. - No completed meta-analysis of breathing and cognition in healthy adults exists — a 2025 systematic-review protocol exists precisely because the literature is unconsolidated. Consequence: a sub-5-minute "focus breath" has no cognitive evidence behind it at all (FR-BR-06 grades accordingly).

Sleep — weak and conflicting; do not headline it. - For: Tsai et al. (2015), Psychophysiology — 20 min at 0.1 Hz pre-sleep reduced sleep-onset latency; n = 28 total [abstract-level]. Laborde et al. (2019) — 30-day evening practice improved subjective sleep quality, n = 64 [secondhand]. - Against: Balban et al. (2023) found no sleep effect after 28 days; the Fincham coherent-breathing RCT (§A2.5) found no change in sleep disturbance in either arm; Lehrer et al. (2020) found sleep among the smallest effects for HRV biofeedback. - Consequence: BR-3 may be framed as wind-down, never as a sleep intervention (NFR-BR-02).

A2.2 What is weak, thin, or contested

Box breathing (4-4-4-4) — insufficient direct evidence, and it has now lost two head-to-heads. Despite being the most recognisable protocol in the culture — and Minimente's current default by inheritance — the protocol-specific base is close to empty, and what exists is unfavourable: - Kasap & Aydin (2025), PLOS ONE 20(11):e0336615 — n = 40, randomised crossover. Box breathing produced significantly higher post-exercise heart rate than 6 breaths/min (164.7 vs 154.8, p < 0.001) and higher perceived exertion (p < 0.001). Context is post-HIIT recovery rather than emotional state, but it is a clean physiological head-to-head and box breathing lost. - Röttger et al. (2021), Applied Psychophysiology and Biofeedback 46(1):19–28 — n = 30, box breathing vs prolonged exhalation during a Stroop task under time pressure and noise. Prolonged exhalation was superior for performance; box breathing produced lower physiological arousal but worse cognitive output. Directly relevant to a desk product: when the user still has to think, the holds cost them. - Balban et al. (2023) — box breathing did not separately beat mindfulness on any outcome; only cyclic sighing did. - Vetter, Kurth & Konigorski (2026), npj Digital Medicine 9:135 — the ASIP study, and the most product-relevant trial in this entire document: N = 76 physicians in residency, series of N-of-1 trials over 4 weeks in the real working day, comparing an 8-minute guided mindfulness/breathing exercise and 6-minute box breathing on daily stress. Population-level effects: 0.18 and 0.07 points on a 10-point scale respectively, with all 95% credible intervals including zero. 10 of 76 met responder criteria; 58% of follow-up respondents nevertheless felt they had benefited. - No meta-analysis and no adequately powered RCT of box breathing for stress or anxiety exists. Its provenance is training culture — Lt. Col. Dave Grossman's On Combat — not a founding study, and product copy must never imply it was "developed by research". - The tactical literature is better than the general one and should be represented fairly: Ibrahim et al. (2023), Military Psychology 36(6):689–700 found 160 seconds of tactical breathing improved marksmanship first-shot score (d = 1.70 — implausibly large, unreplicated, ceiling-affected); Sætrevik et al. (2025), Collabra: Psychology 11(1):144527, preregistered, N = 96 police students, found better rated performance in a simulated critical incident with no associated reduction in self-reported stress; Bouchard et al. (2012), PLoS ONE 7(4):e36169 found large cortisol differences in soldiers but confounded breathing with biofeedback and immersive VR rehearsal. None of this generalises to a desk.

Note also that 4-4-4-4 is 3.75 breaths/min — below the 4.5–6.5 range where the slow-breathing physiology is actually established. The protocol is not merely unevidenced; it sits outside the window the evidence describes.

4-7-8 breathing — weak, and structurally disfavoured here. Popular, sleep-marketed, largely small and uncontrolled studies; recent within-session HRV findings are small and between-group differences minimal. Its seven-count hold is precisely the feature §A5.2 identifies as risky for anxious users. Not recommended for the library.

Alternate-nostril breathing — weak-to-negative, and unusable in context. In the Frontiers in Psychology (2024) systematic review of brief interventions for state anxiety (15:1412928), alternate-nostril breathing yielded g = 0.40 in the wrong direction — worse than control. It also requires a visible hand-to-face posture, failing NFR-PLAT-05 (discretion). Excluded.

Brief breathing for state anxiety generally — mixed, and this deserves emphasis. The same 2024 Frontiers review found breathing-based interventions produced mixed effects on state anxiety, with neither active nor passive breathing significantly outperforming controls overall; deep breathing showed a large effect (g = −1.35) in one direction while alternate-nostril went the other way; heterogeneity was high and publication bias was detected. Progressive muscle relaxation outperformed breathing. This is the most important caveat in the document: the acute, single-session, moment-of-need use case that Minimente's "Right now" surface most needs is less well evidenced than the daily-practice use case the programme delivers.

Energising / alertness breathing — insufficient and actively unsafe. This is the clearest exclusion in the document. - Every technique with a plausible alerting mechanism is hyperventilatory. Hypocapnia causes cerebral vasoconstriction and is documented to produce dizziness, paraesthesia, tetany and syncope. - Fincham et al. (2024), Scientific Reports 14:16893 — n = 200, participant-blinded placebo-controlled RCT of high-ventilation breathwork with retention vs paced breathing: no significant between-group difference on stress (primary), anxiety, depression, wellbeing, sleep or affect, and 9% in each group reported light-headedness, dizziness or tetany. So the energising style is unproven against placebo and carries side effects — and note the 9% appeared in the paced-breathing arm too, which is why FR-BR-21's stop-rule applies to the whole library, not just the excluded protocols. - Fincham, Kartar et al. (2023), Neuroscience & Biobehavioral Reviews 155:105453 documents the mechanism (hypocapnia → alkalosis → altered states), flags unsuitability in cardiovascular disease and psychotic-disorder history, and notes hyperventilation's panicogenic effects. Hyperventilation is also a standard EEG provocation for unmasking seizure activity. - Critical reporting caveat: across the breathwork RCT literature only ~20% of studies reported adverse events at all. "No adverse events were found" is an absence-of-reporting finding, not a safety finding, and Fincham 2023 must never be cited as a safety endorsement. - Cyclic hyperventilation was also the worst-performing arm on state anxiety in Balban 2023. Excluded categorically (NFR-BR-01).

A2.3 Where breathing should not be recommended at all

Acute panic — do not build a mapping. This is the sharpest clinical finding and it cuts against the brief. - Schmidt et al. (2000) dismantling trial (n = 77): breathing retraining added nothing to multicomponent CBT for panic. - Pompoli et al. (2018), Psychological Medicine — component network meta-analysis: interoceptive exposure improves outcomes, while breathing retraining / muscle relaxation components were associated with worse efficacy, though they improved treatment acceptability. - Mechanism concern: breathing used to neutralise feared body sensations functions as a safety behaviour, blocking the corrective learning that drives recovery. Barlow & Craske's Mastery of Your Anxiety and Panic (4th ed.) dropped relaxation training and explicitly reframes breathing as a skill to help clients approach fear, "not intended to reduce fear and anxiety immediately". The Centre for Clinical Interventions' clinician-authored 2023 self-help package When Panic Attacks contains no breathing-retraining module at all, and uses hyperventilation and breath-holding as deliberate exposure tasks. - The counter-position exists and is respectable: Meuret and colleagues' capnometry-assisted respiratory training teaches panic patients to breathe slower and shallower to raise CO₂, with good outcomes — but that is a clinician-supervised, capnometer-instrumented protocol, and its core instruction (breathe less) is the opposite of what every consumer breathing app tells users to do. - Both camps agree "take a deep breath" is wrong advice in panic. - Product consequence (FR-BR-14): Minimente shall not offer a panic-specific breathing protocol, shall make no claim that breathing stops or shortens a panic episode, and shall not present breathing as a way to make frightening body sensations go away.

Trauma histories and dissociation risk. Interoceptive focus, eyes-closed practice and breath-holds are documented triggers; trauma-informed practice orients outward first (SAMHSA TIP 57 grounding; standard trauma-informed guidance on eyes-open practice). Consequence: eyes-open default, an external-anchor protocol in the library, and a visible "this doesn't feel right" exit that routes to grounding rather than to more breathing (FR-BR-15).

Health anxiety. In the Salkovskis & Warwick model, monitoring the body is a maintaining factor, and repeated deep breaths "to check the lungs work" are a documented checking behaviour. Consequence: never instruct the user to scan for whether the technique is "working" (§A5.5 wording rules).

Relaxation-induced anxiety (RIA). Heide & Borkovec (1983/84) documented paradoxical anxiety increase during relaxation; Kim & Newman (2019) tie it to negative-contrast sensitivity in GAD, and peak RIA predicts worse CBT outcome. Reported prevalence estimates in the general literature vary widely (roughly 17–53%). Consequence: the product must normalise this in copy rather than treat it as user failure, and must offer a shorter/eyes-open/grounding alternative rather than "try harder" (FR-BR-16).

A2.4 Contraindications summary (drives §A3 and §A5)

These are the reason the library below is small and hold-free by default. Box Breathing — the product's current default — contains two holds per cycle. That is not a footnote; it is the reason FR-M1-04 is under review (OPEN-BR-01).

A2.5 The central finding: does the evidence support protocol↔state matching?

No. Protocol-to-emotional-state matching is not evidence-supported. On the present literature it would be invented — which the brief explicitly forbids.

The decisive study is one the first draft of this document did not have:

Fincham, Strauss & Cavanagh (2023), Scientific Reports 13:22141. n = 400 (201/199), 10 min/day for 28 days. Intervention: coherent breathing at 5.5 breaths/min. Placebo: an identically structured, guided, nasal, diaphragmatic breathing exercise at 12 breaths/min. Result: no group × time interaction on stress, F(1,377) = 0.089, p = 0.765, ηp² < 0.001 — and nothing on anxiety, depression, wellbeing or sleep either. Both groups improved substantially and equally (time ηp² = 0.161). Credibility and expectancy did not differ between arms. The authors' conclusion: "no measurable effect of coherent breathing over and above a well-designed breathwork placebo."

Read that carefully, because it undercuts more than the resonance rate. In the largest and best-controlled trial in the field, the specific breathing rate — the thing every product in this category sells — added nothing over breathing at more than twice that rate with the same guidance and ritual. A large share of the observed benefit is plausibly expectancy, structure, and taking ten minutes off.

Supporting this reading: - Leyro et al. (2021) — the effect of respiratory therapy shrinks from g = −0.709 against inactive controls to g = −0.279 against active controls, and neither duration nor session count predicts outcome. - Fincham et al. (2023) meta-analysis — no significant differential effects between any subgroups on stress. - Chin et al. (2024), Frontiers in Psychology 15:1412928 — 12 RCTs of brief interventions for state anxiety: breathing was "mixed and inconsistent" (deep breathing g = −1.35; alternate-nostril g = +0.40 favouring control), while embodiment/PMR was the best performer at g = −1.05 and passive attention to the breath outperformed active manipulation of it. I² frequently >75%; publication bias detected. - Balban et al. (2023) — separated from mindfulness only on positive affect and respiratory rate; no differences on trait anxiety, HRV, resting HR or sleep, state-anxiety means heavily overlapping across all four arms (−3.95 / −3.85 / −3.75 / −1.81), and no effect sizes reported anywhere in the paper — p-values only. - Meehan & Shaffer (2024) — the one mechanistic basis for differentiation (exhale ratio) is null in the tightest test.

What is defensible: 1. Slow (~5–6 bpm) beats fast/hyperventilatory. (Fincham subgroups; Balban; Kasap 2025.) 2. A single 5-minute session is sufficient for acute autonomic change, and longer is not better — You et al. (2021), IJERPH 18(23):12478, n = 59, found no RMSSD difference between 5, 10, 15 and 20 minutes. 3. Guided breathing beats doing nothing, and in employees beats a mindfulness comparator for physiological recovery (Chelidoni 2020). 4. Individual response varies enormously and average effects in a real working day are near zero (Vetter 2026: 0.07–0.18 points, CIs spanning zero; 10/76 responders).

What is not defensible: "protocol X for anxiety, protocol Y for focus"; any claim that a particular count or ratio targets a particular emotional state; any claim that the matching engine itself confers benefit over simply guiding slow breathing.

Therefore the recommender in §A4 varies dose, duration and framing across a deliberately small protocol set, and its explanations are worded so as never to imply mechanism-level matching. Recommendation logic in this product is preference, safety and engagement design — not a clinical mechanism, and it is recorded as such here so that nobody downstream writes a marketing claim the evidence cannot carry (§A11, disagreement 1).

---

A3. The breathing library

A3.1 Selection criteria (in order)

1. Contraindication profile — can this be a default for a stranger, unsupervised, at their desk?
2. Evidence quality for the outcome claimed.
3. Distinctness — does it do something the others don't? (Kills near-duplicates.)
4. Silent/discreet operability (NFR-PLAT-02/05).
5. Authoring cost against OPEN-03/OPEN-06/OPEN-15 (§A9).

A3.2 Recommended v1 library — five protocols

FR-BR-01 — The v1 breathing library shall consist of exactly the following five protocols. Additions require the evidence-grading and safety review of §A5.7.

---

BR-1 · Longer Exhale Breathing — the system default

Why it is the default: it sits squarely in the 4.5–6.5 bpm window where the autonomic evidence actually lives, contains no breath-holds, has no meaningful contraindications, works in two minutes at a desk, and matches what public health services publish. The longer exhale is retained because users reliably prefer it (Van Diest 2014) and because it is impossible to over-breathe on — not because superiority over 1:1 is established (Meehan & Shaffer 2024). If Minni prefers a 5-in/5-out pattern, the evidence does not object.

---

BR-2 · Cyclic Sighing (physiological sigh)

---

BR-3 · Resonance Breathing

Honest note, to be repeated in the content: BR-1 and BR-3 are the same protocol family at different doses. They are separated for use-context reasons — a two-minute desk reset versus a deliberate five-minute practice — not because they work by different mechanisms. Given Fincham 2023's placebo null, copy must not imply that BR-3's pace does something BR-1's does not. If OPEN-BR-02 lands on the three-protocol library, BR-3 collapses into BR-1 at a longer duration and nothing of evidential value is lost.

---

BR-4 · Grounded Breathing (external anchor, eyes open)

This protocol simultaneously satisfies the Emergency Tools "Grounding exercise" of FR-ET-03 (§A9).

---

BR-5 · Box Breathing (4-4-4-4) — retained, demoted

Why it is retained at all: it is already authored (Module 1), already familiar to users, already implemented as a pacer in the prototype (§1A.3), and genuinely preferred by some people who like counted structure. Why it is demoted: it is the one protocol in the library whose distinguishing feature is a safety liability with no offsetting evidence, and it lost the only good head-to-head trial. Making it the system default — which is v4's current position by inheritance, not by decision — is not defensible on the evidence.

---

A3.3 Explicitly excluded from the library

FR-BR-02 — The following shall not be implemented in v1, in any tier, and shall not be added without the §A5.7 review:

---

A4. The recommendation engine

A4.1 Need-state vocabulary — extending §7.7, not replacing it

§7.7 FR-NOW-02 established the house vocabulary as plain, non-clinical, self-describing labels (Tense / Racing thoughts / Can't focus / Just checking in). Breathing recommendations use the same states, extended by four. Labels are content data, authored by Minni (FR-NOW-02 precedent).

FR-BR-03 — The breathing recommender shall operate over the following need-state set and shall not introduce any additional state without founder and clinical sign-off:

FR-BR-04 (hard boundary, extends FR-NOW-04) — These are routing inputs, never assessment. No numeric rating, no score, no threshold, no trend, no history view built from need-states, no clinical label anywhere in the UI. The words anxiety, panic, depression, disorder, symptom, diagnosis and treatment shall not appear in breathing-system UI copy.

Stated limitation, per the brief's instruction to say so plainly: the owner's situation list includes acute anxiety and panic symptoms. These cannot be expressed in this vocabulary without introducing clinical labels, and they should not be — both because §2.3/§4.4 forbid it and because §A2.3 shows breathing is the contested-to-wrong intervention for panic. The nearest legitimate expression is too_much, which routes to grounding plus signposting rather than to a "panic protocol". This is a deliberate refusal, recorded at OPEN-BR-03.

A4.2 Activation grouping (internal only, never user-visible)

FR-BR-05 — Each need-state carries an internal activation_class used only to gate hold-containing protocols:

• High activation: tense, wound_up, too_much
• Moderate: racing_thoughts, before_something
• Low / neutral: cant_focus, winding_down, checking_in

Hold-containing protocols (BR-5) shall not be auto-recommended for high-activation states (FR-BR-08). This class is never rendered, never logged against the user, and never exposed via API.

A4.3 The fifteen situations from the brief — honest evidence grading

FR-BR-06 — The following table is the authoritative mapping. Where the grade is weak or none, the system falls back to the default protocol and the explanation must not claim situation-specific evidence (NFR-BR-03).

Read the right-hand column as a whole: eight of fifteen situations collapse onto BR-1 at a differing dose, exactly one situation (#11, recovery) has moderate protocol-relevant evidence in a workplace population, two cannot be modelled at all, and the rest are weak. That is the evidence being reported honestly, not the design being lazy — and it is the strongest possible argument for the small library at §A3 and the near-uniform routing here.

FR-BR-53 (individual-response framing) — Because the only real-workday trial of these exercises (Vetter 2026) found population-level effects near zero with a minority of genuine responders (10 of 76), product copy shall frame breathing as "worth trying for a couple of weeks to see whether it works for you" rather than as an intervention with an expected effect. This is both the honest reading and, per the UX strand, the better engagement design — it sets a testable personal expectation instead of a promise the average user will not experience.

A4.4 First-time users (no history)

FR-BR-07 — For a user with no breathing history, the recommendation shall be produced by this ordered rule list, first match wins:

FR-BR-09 — First-time explanations shall be honest about their basis and shall not simulate personalisation. Indicative wording (Minni authors final):

"A good place to start — slow breathing with a longer out-breath is the most consistently supported pattern, and it takes two minutes."

FR-BR-10 — Exactly one recommendation shall be presented, with a visible, non-prominent "or choose another" affordance opening the full library. Rationale: default effects are large and well evidenced (Jachimowicz et al. 2019, Behavioural Public Policy, d = 0.68); choice overload is conditional but its documented moderators — time pressure, preference uncertainty, low expertise — describe a stressed employee well (Chernev et al. 2015); the escape hatch is what prevents reactance. Note honestly: no trial compares "recommend-one + browse" against "menu-first" in a wellbeing app — this is inference from adjacent evidence, per the UX strand.

FR-BR-11 — No onboarding questionnaire shall be added for breathing. The only input collected before the first exercise is the single need-state tap the user has already given via "Right now" (or nothing at all). Long upfront quizzes are an activation risk with little demonstrated payoff, and §7.7 already supplies the one question worth asking.

A4.5 Returning users (with history)

FR-BR-12 — For a user with history, the engine shall evaluate the following ordered layers. Determinism is absolute: same inputs → same output, always.

Tie-break (FR-BR-12a), applied in order and fully specified so two implementations cannot diverge: (1) higher helped-rate; (2) more completions; (3) most recent success; (4) lower protocol_rank (a fixed integer in content data: BR-1 = 1, BR-2 = 2, BR-3 = 3, BR-4 = 4, BR-5 = 5).

FR-BR-12b (recency decay, deterministic) — Outcomes older than 180 days are excluded from layers 2–4. This is a fixed window, not a decay function; no weighting, no half-life, nothing a user could not reproduce with a calendar.

FR-BR-12c (anti-ossification) — After 8 consecutive completions of the same protocol for the same state, the system shall offer — never impose — one alternative alongside the usual recommendation, with the reason key reason.try_another_optional. Rationale, and its honest limits: habit formation favours consistent repetition of the behaviour (Lally et al. 2010), while identical repeated prompts habituate — the HeartSteps micro-randomised trial (Klasnja et al. 2019) saw suggestion effects decay to near zero within about four weeks. Keeping the behaviour stable while varying the offer is the reconciliation; the specific threshold of 8 is a chosen default, not an evidenced one (OPEN-BR-05).

FR-BR-12d — Explanations are mandatory and literal. Every recommendation carries exactly one reason string, rendered visibly. Indicative wording:

"You said Longer Exhale Breathing helped when you were feeling wound up." "You usually finish Resonance Breathing." "A good place to start — it takes two minutes."

Explanations shall reference only data the user volunteered. This is a substantive design advantage, not a nicety: recommender research shows explanations that reveal invasive or inferred data reduce trust (CHI 2023, "When Recommender Systems Snoop into Social Media, Users Trust Them Less for Health Advice"), while explanation generally increases trust and uptake (Pu & Chen 2007; health-recommender explanation work). Honest limitation: no RCT shows that "because you…" explanations increase completion of wellbeing exercises specifically — the trust finding is extrapolated.

Honest limitation on the entire returning-user layer: a 2024 scoping review of personalisation in mental-health apps (Behaviour & Information Technology) found decision rules and user choice to be the commonest mechanisms and concluded that empirical evidence for personalisation was scarce and inconclusive. History-based recommendation is therefore justified here on grounds of user autonomy, trust and transparency — not on evidence that it improves outcomes. The product should say this internally and should not claim otherwise externally. OPEN-BR-04.

A4.6 Rules as data

FR-BR-17 — All rules in §A4.4 and §A4.5, all need-state labels, all reason strings, all protocol metadata and all thresholds (the 2, the 60%, the 80%, the 180 days, the 8) shall be expressed as versioned content data, editable without an application release, following the FR-PR-02 precedent exactly.

FR-BR-18 — Every recommendation shall log a breathing_recommendation_event (rule_id, need_state, recommended_exercise_id, reason_key, fired_at, accepted bool) — the FR-PR-03 pattern — so rule usefulness is measurable in the pilot. This log contains no free text and is excluded from organisational reporting in any form.

---

A5. Safety

A5.1 Claims discipline

NFR-BR-02 — Breathing content shall make no claim beyond: slow breathing can help you feel steadier, and it is a skill that improves with practice. Prohibited: any claim to treat, cure, prevent or manage a condition; any claim to stop a panic attack; any claim of "proven" effect; any physiological claim the user cannot verify (HRV, vagal tone, "neuroplasticity", cortisol). Competitor practice here is uniformly poor (§A7.6) and is an FTC/ASA-adjacent risk as well as a §2.3 violation.

NFR-BR-03 — Where a recommendation rests on weak or absent situation-specific evidence (§A4.3), the explanation shall be generic ("a good place to start") and shall not imply the protocol was selected for that situation on evidence.

FR-BR-19 (evidence labelling — a genuine differentiator) — Each protocol shall carry a plain-language evidence label rendered in its detail view, drawn from a fixed three-value set: "Well supported", "Some supporting evidence", "Traditional practice, limited evidence". BR-4's label shall be the third value. No competitor reviewed does this (§A7.6).

A5.2 Defaults are shaped by contraindications

FR-BR-20 — No protocol containing a breath-hold shall be a system default for any need-state, for any first-time user, or for any high-activation state. Hold-containing protocols remain available by user choice.

Consequence for FR-M1-04 — a founder decision, presented with both options and a recommendation.

Module 1 currently teaches Box Breathing. It is the first exercise any user ever experiences, at the point of least skill and least trust.

• Option A (recommended). Module 1's taught exercise becomes BR-1 Longer Exhale Breathing. Rationale: no holds, better-evidenced active ingredient, no contraindication load, and it becomes the same object the recommender defaults to everywhere else — one exercise, learned once, reused everywhere (FR-CORE-06). Authoring cost: one exercise script + transcript, not a module rewrite; the surrounding Module 1 copy (welcome, psychoeducation, reflection, daily action) is unaffected in substance.
• Option B. Module 1 keeps Box Breathing; only the recommender defaults change. Authoring cost: zero. Cost incurred elsewhere: the product teaches, as its first act, the one protocol it will then decline to recommend — an incoherence users and reviewers will notice, and the module's exercise carries holds for an audience that includes pregnant users and people with cardiovascular and respiratory conditions.

Recommendation: Option A. Decision: Minni. OPEN-BR-01. Note the interaction with Top Risk 1: Option A adds one script to a project already short ~30% of its copy — but it also removes a script from OPEN-06 (§A9), so the net authoring delta is approximately zero.

A5.3 Universal in-session safety

FR-BR-21 — Every breathing exercise shall display, before the first cycle, a brief non-alarming safety line — shown in full on first-ever use and thereafter available behind a persistent "About this exercise" affordance. Indicative content (Minni authors):

"Breathe gently — never force it. If you feel dizzy, light-headed or uncomfortable at any point, stop and breathe normally. If you're pregnant, or have a heart or breathing condition, keep to gentle breathing and skip any holds."

FR-BR-52 (anti-hyperventilation instruction — binding on every protocol, and the highest value-per-word requirement in this document). Every breathing script shall include, before the first cycle, an explicit instruction not to over-breathe. Indicative wording: "Keep the breaths light and natural — there's no need to fill your lungs."

This is not defensive boilerplate; it is a measured intervention. Szulczewski (2019), Applied Psychophysiology and Biofeedback 44(3):247–256 (n = 46, 10 minutes at 6 bpm) found that paced slow breathing itself induces hypocapnia unless instructed against: end-tidal CO₂ fell 5.21 mmHg in the control group versus 2.70 mmHg in the group told to "avoid excessively deep breathing — breathe shallowly and naturally" (F(1,43) = 4.24, p < .05, η² = 0.09), and hyperventilation symptoms rose significantly more in controls (F(1,37) = 4.63, p < .05).

The implication is blunt: a conventional "breathe deeply and slowly" app prompt is measurably hypocapnia-inducing, and hypocapnia produces exactly the dizziness, tingling and air-hunger that the product is trying to relieve — and that panic-prone users catastrophise. The fix is one sentence of copy and costs nothing. Note also that 9% of participants in the paced-breathing arm of Fincham 2024 reported light-headedness, dizziness or tetany, so this applies to the whole library, not merely to the excluded protocols. NFR-BR-04 already prohibits "take a deep breath"; FR-BR-52 requires the positive instruction that replaces it.

FR-BR-22 (stop-rule) — A Stop control shall be present and reachable at every moment of every exercise, requiring no confirmation, and shall never be styled as a failure. Ending early is a valid outcome (FR-M7-06 precedent).

FR-BR-23 (hold-specific gating) — Before any exercise containing a hold, an additional line shall be shown offering the no-hold alternative in one tap.

A5.4 When the exercise itself goes wrong

FR-BR-16 (relaxation-induced discomfort) — The post-exercise flow shall offer, alongside the neutral outcome capture, a low-prominence "That felt uncomfortable" option. On selection the system shall: 1. Render normalising copy — "That happens to some people, especially at first. It isn't a sign anything is wrong." (grounded in the RIA literature: Heide & Borkovec 1983/84; Kim & Newman 2019); 2. Offer BR-4 Grounded Breathing or a non-breathing Toolkit item — never "try again for longer"; 3. Record a discomfort outcome that feeds the suppression rule BR-R-10.

FR-BR-15 (external-anchor exit) — Every breathing player shall carry a persistent, quiet affordance: "Prefer not to focus on your breathing?" → BR-4. This is the trauma-informed and health-anxiety accommodation, and it costs one link.

A5.5 Wording rules (binding on all breathing content)

NFR-BR-04 — The following are prohibited in any breathing-system copy, notification, or transcript:

Required instead: "as far as is comfortable — no need to force it"; "let the out-breath be a little longer"; "this is a skill; it works best practised when you already feel okay"; "it's normal for this to feel odd at first"; "you can stop at any time." This register matches what NHS and NHS inform actually publish, which is the right benchmark for a non-clinical product.

FR-BR-24 (practise-when-calm framing) — The single most-cited self-guided failure mode is practising only in crisis. checking_in recommendations and programme sessions shall carry practice framing; moment-of-need recommendations shall not lecture.

A5.6 Safeguarding interaction (§8.3) — hooks only, no detection designed here

§8.3 is an unresolved P0 and this addendum does not resolve it. It does, however, create signals that §8.3 will have to consider, and it must not pretend otherwise.

FR-BR-25 (declared signals, not a detector) — The following are safeguarding-relevant and are recorded here so §8.3's designer inherits them explicitly: - Selection of too_much (the most acute need-state); - Repeated breathing/Emergency-Tool use in a short window (the "reaching for it again and again" pattern); - Repeated discomfort outcomes.

FR-BR-26 (what v1 does) — In v1 the breathing system shall: 1. Not implement any threshold, counter, escalation or classifier over these signals. That is §8.3's job and it is undesigned. 2. Ensure the NFR-SAFE-02 "Need urgent help?" signposting is present on every breathing surface, exactly as FR-NOW-05 requires of "Right now". 3. Pin the signposting card above the recommendation whenever need_state = too_much — the NFR-SAFE-07 / FR-SRCH-06 pattern applied to an explicit user statement rather than to scanned content. As with search, responding to what a user deliberately selects is not surveillance of what they write, so this is valid under any NFR-SAFE-04 option Minni chooses. 4. Emit the signals above into safeguarding_events only in the de-identified, rule-id-only form NFR-SAFE-06 already defines, so that when §8.3 is designed the data exists and no retrofit is needed. 5. Never notify an employer, under any circumstance (NFR-SAFE-05).

FR-BR-27 (thinning of the passive signal — carried forward from FR-NOW-06) — Users who rely on breathing rather than programme sessions produce fewer check-ins, thinning NFR-SAFE-03's passive signal. This addendum increases the number of ways to use the product without checking in, and therefore makes always-visible signposting more load-bearing, not less. Recorded as a risk (§A11), not solved here.

A5.7 Adding protocols later

FR-BR-28 — Any addition to the library shall require, before merge: an evidence grade with citations; a completed contraindication row in the §A2.4 format; a hold/no-hold declaration; adaptation copy (NFR-A11Y-04); a reduced-motion pacer definition (§A6.3); and clinical sign-off from the same function that reviews the OPEN-27 crisis vocabulary. "It's popular" is not a criterion.

---

A6. User experience

A6.1 The player

FR-BR-29 — There shall be exactly one breathing player, implemented as a mode of the single exercise runtime (FR-CORE-06), used identically from Module, Toolkit, Emergency Tools, "Right now", Search and Extensions. Presentation overrides (duration preset, framing copy, entry-point-specific follow-up) are permitted; content forks are not (AD-08).

FR-BR-30 (player states) — Prepare → Pace → Complete. Prepare carries the safety line (FR-BR-21), the duration choice, the pacing-mode choice, and the "Prefer not to focus on your breathing?" exit (FR-BR-15). Pace runs the cycle. Complete runs §A6.6.

FR-BR-31 (interruptibility) — Stop is always available (FR-BR-22). A partially completed exercise is recorded as partial with elapsed seconds; it is never rendered as a failure and never affects any streak-like display (UX-09 — there are no streaks).

FR-BR-32 (timing anchor) — Cycle timing shall be anchored to a monotonic clock and shall remain correct across backgrounding, screen lock and tab suspension, per AD-07's reasoning. A breathing pacer that drifts is worse than none.

A6.2 Pacing modes

FR-BR-33 — The player shall support four pacing modes, user-selectable and remembered per user:

Modes are combinable except that audio is never enabled without an explicit user action in-session.

A6.3 The pacer, and the reduced-motion requirement

An expanding/contracting full-screen shape is the obvious implementation and the obvious NFR-A11Y-05 violation: large-area scaling motion is a known vestibular trigger, and this is exactly where breathing products fail.

NFR-BR-05 (binding) — The player shall detect prefers-reduced-motion and, when set, shall render the visual_reduced pacer by default, without the user having to find a setting. The visual_reduced pacer shall: - Use no scaling, no translation, no rotation, and no continuous motion of any element; - Convey phase by explicit text ("Breathe in" / "Hold" / "Breathe out") plus a discrete numeric countdown that changes at most once per second; - Use a linear determinate progress indicator that updates in discrete steps, or a static ring with a stepped fill; - Never rely on colour alone to signal phase (NFR-A11Y-06) — phase is carried by text in every mode; - Be independently selectable by any user, whether or not the OS preference is set.

NFR-BR-05a — The animated pacer's motion shall additionally be bounded: no full-viewport scaling, no parallax, no particle effects, and a maximum scale ratio of 1:1.6 between the smallest and largest state.

NFR-BR-05b (screen reader) — In every mode, phase transitions shall be announced via a polite live region, with announcements limited to the phase word and count. Screen-reader users shall be able to complete an exercise with the pacer alone (NFR-A11Y-01).

NFR-BR-05c — All controls meet the 44×44pt target minimum (NFR-A11Y-07) and remain functional at 200% text scaling (NFR-A11Y-05).

A6.4 Silent, desk-based operation

NFR-BR-06 — Every exercise in the library shall be completable end-to-end with the device muted and no headphones, with no loss of guidance. Audio is an enhancement; the transcript and the visual/text pacer are the primary channel (NFR-A11Y-03, UX-05). No autoplay, ever (FR-M1-04 precedent).

NFR-BR-07 (discretion — NFR-PLAT-05) — The breathing player shall be visually quiet: no full-screen colour washes, no large "MINDFULNESS"/"MENTAL HEALTH" typography, no celebratory animation, and no branding on the pacer screen that identifies the app's purpose from a metre away. It should be plausibly mistakable for a timer. The excluded protocols help here too: nothing in the library requires an audible exhale, a visible hand-to-face posture, or closed eyes.

FR-BR-34 (nose-exhale variant) — BR-2's mouth exhale is mildly audible; a nose-exhale variant shall be offered and shall be the default when the user has selected a "quiet surroundings" preference.

A6.5 Haptics

FR-BR-35 — Where the platform supports vibration, the player shall offer haptic phase marking with intensity options Off / Light / Distinct, following the pattern Apple's Watch Breathe session established (None / Minimal / Prominent) — which remains the best-executed silent pacing in the market.

FR-BR-36 (platform reality, stated so nobody plans around a fantasy) — The Web Vibration API is not implemented in Safari on any Apple platform. A PWA (AD-02) therefore cannot deliver haptic pacing on iPhone or iPad without a native wrapper. Haptics shall be feature-detected and the mode hidden where unavailable; haptic pacing shall not be treated as a cross-platform capability, and no content shall instruct a user to "feel the taps" unless the mode is active. This is a genuine input to the OPEN-24 stack decision and should be raised in that discussion.

A6.6 Completion flow and outcome capture

FR-BR-37 — On completion (or early stop), the player shall present: a factual, unembellished acknowledgement; the optional outcome capture below; and at most two onward actions (typically Add to my Toolkit per FR-CORE-07/FR-STD-07, and Done).

FR-BR-38 (outcome capture — neutral, one tap, skippable) — The system may ask one question, with three options and a skip:

"How do you feel now?" → A bit steadier / About the same / That felt uncomfortable

Design constraints, each with a reason: - Neutral wording, not "Did it help?" — demand characteristics inflate agreement with the app's evident preferred answer (Dell et al., CHI 2012, "Yours is better!"; self-reported adherence to digital interventions is inflated generally). Expect a positive skew regardless and treat the data as within-user relative signal only, never as an efficacy measure. - Always skippable; never blocks anything (UX-07, FR-STD-04). - Sampled, not universal: asked on every completion for the first 5 completions, then at most every third completion. EMA research consistently shows compliance falls as prompt frequency rises; once-daily prompting substantially outperforms higher rates. - The third option is the discomfort path (FR-BR-16), deliberately placed here so that reporting a bad experience is as easy as reporting a good one.

FR-BR-39 — Outcome responses feed only the §A4.5 recommender and the user's own view. They are never aggregated into an efficacy claim, never shown to an employer (NFR-PRIV-06, FR-TK-08), and never presented back to the user as a trend line or score (FR-BR-04, FR-PROG-04's non-diagnostic framing).

FR-BR-39a (the telemetry will under-detect — do not treat it as an outcome measure) — Across Plans (2019), Chelidoni (2020) and Sætrevik (2025), physiological and performance benefits appeared repeatedly without matching self-reported improvement; in Plans 2019 subjective stress recovered in the control group too. The product's own one-tap self-report is therefore biased in both directions at once — inflated by demand characteristics, deflated by genuine insensitivity to physiological change. It is adequate for its actual job (ranking protocols within one user) and inadequate for any other. It shall not be used as a pilot success metric (§4.4 already relies on the programme's own check-ins) and shall not appear in any efficacy claim.

A6.7 Favourites, recents, progress

• FR-BR-40 — Any protocol may be favourited from the player or the library; favourites are Toolkit items of type built_in_exercise (FR-TK-03), idempotent (FR-TK-06), and directly runnable (FR-TK-05). No parallel favourites store shall be created.
• FR-BR-41 — A "Recently used" list of at most 5 protocols shall appear in the breathing library view, ordered by recency. Local, ephemeral, non-scored.
• FR-BR-42 — Progress display is factual only: protocols practised, total minutes if shown at all. No streaks, no badges, no chains, no "you've missed…" (UX-09, §2.3 — binding). A breathing feature is the most tempting place in the product to add a streak; it is prohibited here as everywhere.

---

A7. Integration — breathing as a subsystem, not a feature

AD-BR-01 — Breathing is a specialisation of FR-CORE-06's exercise object, not a parallel system: protocols are rows in exercises, they run in the one exercise runtime, and every surface holds references. There is no "breathing service", no separate content pipeline, and no second player.

A7.1 Programme / Exercise Library

• Module 1 teaches one protocol (OPEN-BR-01 decides which) and offers it to the Toolkit (FR-CORE-07).
• Module 11's "Calm Breathing" branch (FR-M11-02) shall reference BR-1, not a fourth copy of a breathing script — removing one of the four missing branch scripts from OPEN-02's authoring list.
• Extension modules may reference any library protocol without new authoring.

A7.2 "Right now" (§7.7)

• The breathing recommender is the breathing half of FR-NOW-03's routing table, extended with the four new states (FR-BR-03). One routing mechanism, one table format, one set of content keys.
• FR-NOW-03's stated preference — the user's own saved tool first, else a sensible default — is implemented by §A4.5's layers 2–4 (history) falling through to §A4.4 (evidence defaults). These are the same policy, expressed once.
• FR-NOW-04's hard boundary applies unchanged and is restated as FR-BR-04.

A7.3 Toolkit (§7.1)

• Protocols accrue via FR-CORE-07/FR-STD-07 like any exercise; runnable in place (FR-TK-05); private (FR-TK-08).
• FR-BR-43 — A Toolkit item may carry a user-chosen need_state association ("this one's for when I'm wound up"). This is the only user-authored personalisation input in the system, it is optional, and it feeds §A4.5 layer 2 as a tie-break preference, never as an override of the safety layers.

A7.4 Emergency Tools (§7.2)

• FR-BR-44 — "Breathing reset" (FR-ET-03) resolves to BR-1 with a 90-second emergency preset; "Grounding exercise" resolves to BR-4. Per FR-ET-08 these are the same objects, not copies.
• Emergency-context presets are shorter by default and skip the practice framing.
• FR-ET-07's "Need urgent help?" adjacency applies, and FR-BR-26 pins it for too_much.

A7.5 Search (§7.8)

• FR-BR-45 — Protocols join the content index with their tags ("breathing", "longer exhale", "box breathing", "can't switch off", "wound up"). Everyday-language synonyms are authored into the OPEN-27 vocabulary — this addendum adds terms to that artefact, it does not create a second one.
• Evidence labels are indexed, so a user searching "box breathing" reaches BR-5 and sees its honest label rather than being quietly redirected. Do not hide a protocol a user explicitly searched for; demotion in recommendation is legitimate, suppression in retrieval is paternalistic and erodes trust.
• FR-SRCH-06 / NFR-SAFE-07 crisis routing is unchanged and takes precedence over any breathing result.

A7.6 What the market does, and the gap this occupies

From the product-landscape strand (Breathwrk, Othership, Calm, Headspace, Breathly, Balance, Oak, iBreathe, Paced Breathing, Apple Watch Breathe):

• The market is a barbell: content platforms with 50–500+ exercises monetised by volume and retention mechanics, and minimalist utilities with 1–10 exercises that users praise precisely for restraint. Nothing occupies the middle — a small, curated, intelligently-recommended library.
• Nobody explains why. Choice is by browsing (Breathly, iBreathe, Oak), goal category (Breathwrk's Calming/Sleep/Energizing/Perform/Health; Othership's UP/DOWN/ALL AROUND; Balance's Relax/Energize/Concentrate), or silent adaptation (Balance — which adapts but hides its reasoning). No product surfaces the reason for a recommendation.
• Evidence labelling is uniformly weak, ranging from vague ("informed by modern science") to inflated ("neuroscience based", "long-term neuroplasticity") to survey-stat marketing ("95% of users report less stress"). None cites studies at the point of the exercise; none grades its own evidence; none says "this one is traditional practice with limited trials".
• Safety framing is near-absent in the products shipping the riskiest content: high-ventilation protocols reach consumers with minimal in-flow contraindication warnings, and user reviews of at least one such app report adverse physical effects.
• Silent/haptic operation is solved at the utility tier and ignored by the platforms. Apple's three-level haptic pacing is the benchmark; Calm buries haptics in profile settings; Othership effectively requires audio and eyes closed — unusable in an office.
• Accessibility is an open field. No major breathing product documents a reduced-motion alternative to its pacer animation.

The gap Minimente can occupy: small honest library · explained deterministic recommendation · graded evidence · safety-first defaults · silent by design · no gamification. Every one of those is already a Minimente constraint rather than a new investment — which is the strongest argument for building this at all.

---

A8. Technical specification

Extends the v4 architecture only where necessary; redesigns nothing.

A8.1 No new services

AD-BR-02 — No new backend service. The Content Service gains breathing protocols and recommendation rules as content; the Programme & Response Service evaluates the rules (exactly as it already evaluates §7.3 and §8.3 rules) and stores outcomes; the client gains a pacer module inside the existing exercise runtime. §9.1's diagram is unchanged.

AD-BR-03 — Rule evaluation is deterministic and side-effect-free, and must be implementable identically on client and server. The client may evaluate locally for offline use (NFR-PLAT-04); the server evaluation is authoritative on conflict. Same rule data, same result — this is a testable property and should have a test (§A8.5).

A8.2 Schema — additive only

Extends §7.4. No existing table or column is redefined.

exercises                                  -- EXISTING (v4). Two columns added.
  ...
  breathing_config JSONB NULL              -- NEW: present iff this exercise is a breathing protocol
                                           --   { pattern: [{phase:'inhale',   seconds:4},
                                           --               {phase:'exhale',   seconds:6}],
                                           --     cycles_or_duration: 'duration',
                                           --     durations_s: [60,120,300],
                                           --     default_duration_s: 120,
                                           --     has_breath_hold: false,
                                           --     max_hold_s: 0,
                                           --     protocol_rank: 1,
                                           --     evidence_grade: 'some_support',
                                           --     nose_exhale_variant: true,
                                           --     reduced_motion_labels: {inhale:'…', exhale:'…'} }
  contraindication_flags TEXT[] NULL       -- NEW: e.g. {'holds','pregnancy_caution','cardio_caution'}
                                           --   drives FR-BR-08/20/23 gating

breathing_recommendation_rules             -- NEW: content data, versioned, editable without release
  id, rule_id ('BR-R-04'), order_index,
  condition JSONB,                         -- declarative; no expressions, no code
  recommended_exercise_id, duration_s,
  reason_key, active BOOL, version, published_at

exercise_outcomes                          -- NEW: user data. The recommender's only learning input.
  id, user_id, exercise_id,
  need_state TEXT NULL,                    -- enum from FR-BR-03; never free text
  entry_point TEXT,                        -- 'module'|'toolkit'|'right_now'|'emergency'|'search'
  started_at, completed BOOL, elapsed_seconds,
  outcome TEXT NULL,                       -- 'steadier'|'same'|'discomfort'|NULL(skipped)
  pacing_mode TEXT                         -- 'visual'|'visual_reduced'|'haptic'|'audio'
  -- NO free text column. Deliberate: see FR-BR-46.

breathing_recommendation_events            -- NEW: FR-PR-03 pattern, for pilot measurement
  id, user_id, rule_id, need_state, recommended_exercise_id,
  reason_key, fired_at, accepted BOOL NULL

FR-BR-46 — exercise_outcomes shall contain no free-text column, and none shall be added. Free text would make the recommender's input special-category data under FR-DATA-02 and would create a second unscanned disclosure channel interacting with the undesigned NFR-SAFE-04 decision. A three-value enum is sufficient for every rule in §A4.5.

FR-BR-47 — Emergency-tool breathing usage continues to write emergency_tool_events (FR-ET-04) and exercise_outcomes. The two are joinable on exercise_id but serve different purposes; neither is duplicated into the other.

A8.3 Content model

FR-BR-48 — A breathing protocol is an exercises row with breathing_config populated, its own copy_status (FR-CORE-03 — a protocol with placeholder copy is never recommended, never surfaced in search, and never offered in the library), its own locale variants (NFR-I18N-02), and its own adaptation_copy_key (NFR-A11Y-04).

FR-BR-49 — Rule content is validated at publish time. The Content Service shall reject a rule set that: references a non-existent or non-final exercise; recommends a has_breath_hold protocol for a high-activation state (FR-BR-08); leaves any need-state without a terminal fallback; or contains an unreferenced reason_key. These are CI checks against content, matching the §11 item-18 precedent.

A8.4 Privacy

• NFR-BR-09 — exercise_outcomes and breathing_recommendation_events are user data, subject to §8.2 in full: EU residency (NFR-PRIV-08), RLS isolation (NFR-SEC-04, AD-04), export and hard delete (NFR-PRIV-07, FR-DATA-04).
• NFR-BR-10 — Neither table shall ever enter organisational reporting, in any aggregated, summarised or sampled form (NFR-PRIV-06, FR-TK-08). Breathing usage may contribute only to the existing undifferentiated "sustained use" aggregate (FR-POST-05) — an employer must not be able to learn that breathing specifically is what a cohort reaches for, since that is closer to an inference about the workforce's state than to a usage metric.
• NFR-BR-11 (retention) — Outcomes are retained while the account is active and for 12 months thereafter, consistent with NFR-PRIV-09's recommendation. Because FR-BR-12b already ignores outcomes older than 180 days, longer retention buys the product nothing and should not be requested.
• NFR-BR-12 — Need-state selections follow the right_now_events precedent exactly: routing signal only, excluded from org reporting, never scored, never trended (FR-NOW-04).

A8.5 Testability

FR-BR-50 — The rule engine shall ship with a golden-file test suite: a fixed set of (history, need-state) inputs mapped to expected (exercise, duration, reason_key) outputs, versioned alongside the rule content. Any rule edit that changes an output must change the golden file in the same commit. This is what makes "deterministic and explainable" an enforced property rather than an aspiration.

---

A9. Content and authoring burden — the honest accounting

Content is the project's #1 risk (Top Risk 1): ~30% of the programme has no copy (OPEN-03), the Emergency Tools have no scripts at all (OPEN-06), and audio is unplanned (OPEN-15). A breathing library is not free, and this section refuses to pretend otherwise.

A9.1 What each protocol costs to author

Per protocol: (1) exercise script; (2) full transcript — required whether or not audio is ever recorded (NFR-A11Y-03, UX-05); (3) adaptation copy (NFR-A11Y-04); (4) reduced-motion phase labels; (5) short description + evidence label; (6) contraindication line if hold-containing. Roughly half a day of authoring per protocol for someone who knows the material — plus review.

A9.2 Net position for the five-protocol library

Offsetting savings, which are real: - OPEN-06 shrinks. "Breathing reset" and "Grounding exercise" are two of the five (or four) unwritten Emergency Tool scripts. Both are satisfied by BR-1 and BR-4. OPEN-06's authoring count drops by two. - OPEN-02 shrinks by one. Module 11's Calm Breathing branch references BR-1 instead of needing a fifth breathing script. - If OPEN-BR-01 lands on Option A, Module 1's script is replaced rather than added — cost neutral against the count above.

Net: approximately +2 to +3 new scripts against the pre-existing plan, plus one safety-copy set and a set of short strings. This is the smallest honest number for a five-protocol library, and it is achievable.

A9.3 Would a smaller library be the better product? — asked and answered

Yes, materially — and it is worth putting on the table explicitly. A three-protocol library (BR-1 Longer Exhale, BR-4 Grounded, BR-5 Box retained as the already-authored familiar option) would: - lose almost nothing on the evidence, since BR-3 is BR-1 at a longer dose and BR-2's advantage is modest and single-source; - cost one new script instead of four; - still satisfy every rule in §A4.4 with a trivial substitution (BR-2 → BR-1 at 60 s; BR-3 → BR-1 at 5 min).

The case for five rather than three is: BR-2's 60-second option is genuinely the best fit for the "I have one minute at my desk" case that the workplace positioning depends on; BR-3's longer dose is where the sleep and training-effect evidence actually lives; and Oak's three-exercise precedent notwithstanding, five is still far below every competitor.

Recommendation: ship five, but sequence them — BR-1, BR-4 and BR-5 at launch (BR-5 already exists), BR-2 and BR-3 as the first content increment. This lets the system ship with one genuinely new script. OPEN-BR-02.

A9.4 Audio (OPEN-15)

FR-BR-51 — The breathing system shall ship text-and-pacer-only if audio is unresolved. Every protocol is fully usable with no audio — that is a requirement (NFR-BR-06), not a degradation. Breathing is the strongest candidate in the entire product for shipping without audio, because the pacer is the guidance. This should relieve, not increase, pressure on OPEN-15.

---

A10. Open questions

Format per §10. P0 = blocks launch · P1 = blocks build of the affected area · P2 = needs a decision before pilot.

---

A11. Risks, and where this addendum disagrees with the brief

A11.1 New risks introduced

1. Acute-use evidence is weaker than daily-practice evidence. The 2024 Frontiers review found brief breathing produced mixed effects on state anxiety, with high heterogeneity and detected publication bias. The product's most prominent breathing surface ("Right now") is therefore its least evidenced. Mitigation: modest claims (NFR-BR-02), and grounding as an alternative that never over-promises.
2. More ways to use the product without checking in. Every surface this addendum adds is a route that produces no daily_checkin, thinning NFR-SAFE-03's passive signal further (FR-BR-27). §8.3 must be designed knowing this.
3. A new user-data class. exercise_outcomes is new personal data in a product whose central promise is privacy. Constrained hard (§A8.4) and deliberately free-text-free (FR-BR-46) — but it exists, and the DPIA (NFR-PRIV-01) must cover it.
4. The recommender can be wrong in a visible way. Deterministic and explained means a bad recommendation is legible and attributable. That is the correct trade against an opaque one, but it raises the stakes on the rule content.

A11.2 Where the brief is wrong, unsupported, or in tension with the PRD

Stated directly, because the brief asked for disagreement rather than compliance.

1. "Intelligent matching of protocols to emotional states" is not supported by evidence — and the failure is deeper than expected. This is the brief's central premise and it does not survive the literature. The decisive result is Fincham 2023 (Sci Rep 13:22141): in a 400-person placebo-controlled RCT, coherent breathing at 5.5 bpm was no better than a guided placebo at 12 bpm on stress, anxiety, depression, wellbeing or sleep. If the rate — the most-sold parameter in this entire category — adds nothing over a credible ritual, then a matrix mapping fifteen emotional states to distinct patterns is not a refinement of the evidence, it is fiction. Supporting this: effects shrink to g = −0.279 against active controls (Leyro 2021); no subgroup differences on stress (Fincham 2023 meta-analysis); the exhale-ratio mechanism is null in its tightest test (Meehan & Shaffer 2024); and passive attention to the breath outperformed active manipulation of it (Chin 2024). The system specified here therefore varies dose, duration and framing across five protocols, eight of the fifteen situations converge on the same protocol, and the recommender is documented internally as preference/safety/engagement design rather than a clinical mechanism. The brief forbids inventing recommendations; applied consistently, that principle rules against the brief's own framing.

2. Two of the fifteen situations should not be modelled at all — and one of them is dangerous to model. "Panic symptoms" fails twice: it is a clinical label the PRD forbids (§2.3, §4.4), and the clinical evidence is that breathing retraining is at best non-essential and at worst counterproductive for panic (Schmidt 2000; Pompoli 2018, where breathing/relaxation components were associated with worse efficacy; Barlow & Craske's own manual dropped and reframed it; CCI's current clinician-authored package omits it entirely). Building a "panic breathing" route would put Minimente on the wrong side of the modern evidence while breaking its own non-diagnostic constraint. Declined. "Low energy" fails differently: there is no safe, evidenced protocol, and every candidate sits in the high-ventilation family the product must exclude. The honest answer to "I have no energy" is not a breathing exercise, and the system says so (FR-BR-13).

3. Box Breathing as the incumbent default is not defensible, and the case is stronger than a safety argument alone. Its distinguishing feature — two breath-holds per cycle — carries a contraindication load (pregnancy, cardiovascular, respiratory) and a documented air-hunger risk for anxious users, with no offsetting benefit evidence. Beyond that, it has now lost every head-to-head located: higher heart rate and higher perceived exertion than 6 bpm (Kasap & Aydin 2025); worse cognitive performance than prolonged exhalation under load (Röttger 2021); no separation from mindfulness in Balban 2023; and 0.07 points on a 10-point scale with a credible interval spanning zero in the only real-workday trial (Vetter 2026). At 3.75 bpm it also sits below the rate window where the physiology is established, and its provenance is On Combat, not a study. v4 made it the default by inheritance from a prototype, not by decision. Retaining it as a choice is right — people genuinely like the counted structure; retaining it as the default, and teaching it as the user's first-ever exercise, is not.

4. "The system should learn over time" is a product-preference decision, not an evidence-based one. The brief presents previous-success learning as self-evidently right. The personalisation literature in mental-health apps is explicitly "scarce and inconclusive". Building it is defensible on autonomy, trust and explanation-quality grounds — which are real — but the document should not, and does not, claim it improves outcomes. A leaner v1 (evidence defaults + favourites only) is a legitimate alternative and is recorded as OPEN-BR-04 rather than being quietly dismissed.

5. The brief under-weights how much of this is content cost. "Research and recommend the complete library" is phrased as a research task; it is mostly an authoring commitment against a project that is already ~30% short of copy, has zero Emergency Tool scripts, and has no audio plan. This addendum's answer — five protocols, phased, with two of them retiring existing OPEN-06 obligations — is deliberately smaller than the brief implies, and §A9.3 argues openly that three might be better still.

6. Silent-first and reduced-motion are not accessibility line-items here; they are the primary design path. A breathing pacer is the single most likely place in the product to violate NFR-A11Y-05, and an office breathing player is the single most likely place to violate NFR-PLAT-02/05. The brief lists "visual pacing, audio guidance, silent mode, vibration support" in that order; the correct order for this product is text/visual-reduced first, visual second, haptic third (where the platform even permits it), audio last and never by default.

7. The brief's implicit promise — that the right exercise produces a reliable effect — is not one the product can keep, and the copy must be built around that. In the only trial conducted in a real working day (Vetter 2026, N = 76 physicians, N-of-1 design), average daily-stress effects were 0.07–0.18 points on a 10-point scale with credible intervals including zero, while 58% of respondents nonetheless felt they had benefited and 10 of 76 met responder criteria. The honest product framing is individual response — "try it for a fortnight and see whether it works for you" (FR-BR-53) — not an expected benefit. This is a better engagement design as well as a truer one, but it is not what the brief assumes.

8. One thing in the brief is more right than it may have realised. "Explainable, deterministic, non-clinical" is not merely a constraint imposed by §4.3's AI exclusion — the product research found that no competitor explains its recommendations and none grades its own evidence. The honest-evidence-label plus stated-reason combination is the clearest differentiation available to Minimente in this category, and it costs almost nothing to build because the PRD's existing constraints already force most of it.

---

Appendix A.2 — Principal sources

Evidence base — reviews and meta-analyses. Fincham, Strauss, Montero-Marin & Cavanagh (2023), Scientific Reports 13:432 — breathwork meta-analysis. Laborde et al. (2022), Neuroscience & Biobehavioral Reviews 138:104711 — slow breathing and HRV, 223 studies. Shao, Man & Lee (2024), Mindfulness — slow-paced breathing, cardiovascular and emotion functions, 31 studies. Leyro et al. (2021), Clinical Psychology Review 84:101980 — respiratory therapy for anxiety, 40 RCTs. Zaccaro et al. (2018), Frontiers in Human Neuroscience 12:353. Goessl, Curtiss & Hofmann (2017), Psychological Medicine 47:2578–2586. Vann-Adibe et al. (2025), Applied Psychophysiology and Biofeedback — remote HRVB, 18 RCTs. Lehrer et al. (2020), Applied Psychophysiology and Biofeedback. Chin et al. (2024), Frontiers in Psychology 15:1412928 — brief interventions for state anxiety. Hopper et al. (2019), JBI 17(9):1855–1876 [abstract-level]. Fincham, Kartar et al. (2023), Neuroscience & Biobehavioral Reviews 155:105453 — high-ventilation breathwork [partially verified; full text 403-blocked].

Evidence base — primary trials. Fincham, Strauss & Cavanagh (2023), Scientific Reports 13:22141 — coherent breathing vs placebo, n = 400, null (the pivotal result of §A2.5). Fincham et al. (2024), Scientific Reports 14:16893 — high-ventilation breathwork vs paced breathing, n = 200, null with 9% adverse sensations in both arms. Balban et al. (2023), Cell Reports Medicine 4(1):100895. Van Diest et al. (2014), Applied Psychophysiology and Biofeedback 39(3-4):171–180. Meehan & Shaffer (2024), Applied Psychophysiology and Biofeedback 49(3):407–417 — exhale-ratio null. Szulczewski (2019), Applied Psychophysiology and Biofeedback 44(3):247–256 — anti-hyperventilation instruction. You et al. (2021), IJERPH 18(23):12478 — 5/10/15/20-minute dose-response. Chelidoni et al. (2020), JMIR mHealth and uHealth 8(10):e19412 — employees. Plans et al. (2019), JMIR Formative Research 3(1):e12227. Vetter, Kurth & Konigorski (2026), npj Digital Medicine 9:135 — ASIP N-of-1 trials in physicians. Kasap & Aydin (2025), PLOS ONE 20(11):e0336615 — box breathing vs 6 bpm. Röttger et al. (2021), Applied Psychophysiology and Biofeedback 46(1):19–28. Bouchard et al. (2012), PLoS ONE 7(4):e36169. Ibrahim et al. (2023), Military Psychology 36(6):689–700. Sætrevik et al. (2025), Collabra: Psychology 11(1):144527. Ma et al. (2017), Frontiers in Psychology 8:874. Laborde et al. (2022), Journal of Psychophysiology 36(1):13–27. Blaser, Weymar & Wendt (2023), Frontiers in Psychology 14:1292983. Wells et al. (2012), PLoS ONE 7(10):e46597. Tsai et al. (2015), Psychophysiology [abstract-level]. Albulescu et al. (2022), PLOS ONE 17(8):e0272460 — micro-break meta-analysis (breaks generally, not breathing).

Clinical practice. Schmidt et al. (2000) — CBT dismantling for panic. Pompoli et al. (2018), Psychological Medicine — component network meta-analysis. Barlow & Craske, Mastery of Your Anxiety and Panic, 4th ed. Meuret et al. — capnometry-assisted respiratory training. Heide & Borkovec (1983/84); Kim & Newman (2019), Journal of Affective Disorders — relaxation-induced anxiety and contrast avoidance. Salkovskis & Warwick — CBT model of health anxiety. SAMHSA TIP 57 — grounding. Centre for Clinical Interventions, When Panic Attacks (2023). NHS "Breathing exercises for stress"; NHS inform controlled and belly breathing.

Product landscape. Breathwrk, Othership, Calm, Headspace, Breathly, Balance, Oak, iBreathe, Paced Breathing, Apple Watch Breathe/Mindfulness — vendor sites, App Store/Play listings, support documentation and user reviews, July 2026. Apple's reduced-motion evaluation criteria for App Store accessibility labels. Web Vibration API browser-support data (Safari: not implemented on any Apple platform).

UX and personalisation. Jachimowicz et al. (2019), Behavioural Public Policy — default effects meta-analysis. Scheibehenne, Greifeneder & Todd (2010), Journal of Consumer Research; Chernev, Böckenholt & Goodman (2015), Journal of Consumer Psychology — choice overload and its moderators. Nahum-Shani et al. (2016/2018) — JITAI framework. Klasnja et al. (2019), Annals of Behavioral Medicine — HeartSteps habituation. Lally et al. (2010), European Journal of Social Psychology — habit formation. Pu & Chen (2007) — explanation interfaces and trust; CHI 2023 — recommender "snooping" and health-advice trust. Dell et al. (2012), CHI — "Yours is better!" demand characteristics. Behaviour & Information Technology (2024) — personalisation in mental-health apps scoping review. JMIR (2021) — EMA compliance meta-analysis.

Research-strand provenance. All five commissioned strands — scientific literature, clinical practice, performance & human factors, product landscape, UX & personalisation — completed and are reflected here. The literature and performance strands reported late and corrected three conclusions in the first draft (§A0.3): the resonance-rate rationale, the box-breathing head-to-head record, and the status of exhale emphasis. Sources verified only at abstract or record level are flagged [abstract-level] or [secondhand] inline; OPEN-BR-08 carries a single confirmatory clinical pass before build.