Since coming out of stealth, we have received a lot of questions — from scientists, clinicians, wearable industry veterans, and women who want to understand what Clair actually does and how it differs from what already exists. Many of these questions are good. Some reflect genuine skepticism about whether non-invasive hormone monitoring is possible. We welcome that skepticism. It makes the science better.

This post addresses the most common questions directly. We are going to be specific about what we can do, what we cannot do yet, and where the open scientific questions are.

The Science

How does Clair track hormone levels without a blood draw?

Hormones are not abstract. They are physical molecules that bind to receptors across your body and produce measurable physiological effects — continuously, across multiple organ systems, around the clock. Clair's approach is built on this foundational biology.

When progesterone rises after ovulation, it binds to receptors in the hypothalamus and shifts your thermoregulatory set point upward by 0.3–0.5°C. It simultaneously increases resting heart rate by approximately 3.5 BPM, reduces parasympathetic-mediated heart rate variability (RMSSD) by roughly 12%^[1][2], and alters body composition through fluid retention measurable via bioimpedance. Estrogen, rising during the follicular phase, promotes vasodilation — shifting arterial stiffness, pulse wave morphology, and the spectral characteristics of photoplethysmographic signals^[8]. It modulates autonomic tone in the opposite direction from progesterone, increasing vagal HRV^[2]. The LH surge that triggers ovulation produces a cascade of rapid, concurrent physiological changes detectable across cardiovascular, thermoregulatory, and autonomic channels within a narrow time window.

These are not subtle effects if you are measuring with sufficient precision across enough physiological domains. They are well-documented in peer-reviewed literature going back decades. What has been missing is a wearable platform purpose-built to capture them all simultaneously — and machine learning models trained on paired wearable-and-hormone data to decode the combined pattern.

That is what Clair is. We built a multi-modal sensor array that captures over 130 mechanistic biomarkers spanning cardiovascular, thermoregulatory, autonomic, electrodermal, body composition, respiratory, acoustic, sleep, and activity domains. Our models learn the mapping between these multi-system physiological signatures and the underlying hormonal state — trained against clinical-grade hormone ground truth, not just self-reported cycle data. The result is continuous hormonal inference from signals the body is already producing.

What is the difference between monitoring, tracking, and measuring?

This distinction matters, and it is less clear-cut across medicine than most people realize.

Measuring means directly quantifying a specific analyte. A blood draw analyzed by LC-MS/MS that returns "estradiol: 150 pg/mL" is a direct measurement. It detects the hormone molecule itself via mass spectrometry.

Monitoring means continuously observing a system's state through available signals — which may or may not involve direct analyte measurement. In clinical practice, most continuous monitoring is indirect. A continuous glucose monitor (CGM) does not measure blood glucose. It measures glucose in interstitial fluid via an enzymatic reaction on a subcutaneous sensor, then infers blood glucose from that proxy — with an inherent 5–20 minute lag^[4][5]. A pulse oximeter does not measure arterial oxygen saturation. It measures the differential absorption of two wavelengths of light through tissue and infers SpO2 from the ratio^[6]. A fetal heart rate monitor does not measure fetal cardiac function directly — it detects Doppler-shifted ultrasound reflections and infers heart rate from the signal.

Yet CGMs, pulse oximeters, and fetal monitors are all standard of care. They work not because they measure directly, but because the inference is well-validated, clinically useful, and robust enough for the intended application.

Tracking means logging discrete data points over time — typically at low frequency, often with user input. Period tracking apps that ask you to log your start date are tracking. BBT methods that require a single morning reading are tracking.

Clair is a continuous hormone monitor. It continuously observes your hormonal state through multi-modal physiological signals — analogous to how a CGM continuously monitors glucose through interstitial fluid. Neither measures the target analyte directly. Both infer it from correlated signals at high frequency, validated against ground-truth measurements. The question, for any monitoring technology, is not "is it direct?" but "how well-validated is the inference, and is it clinically useful?" That is what our validation data and clinical study program are designed to establish.

Isn't this just inference? How is that valid?

Yes, it is inference — and so is virtually every hormone test that exists.

Urine-based ovulation and fertility tests do not measure estradiol or progesterone. They measure E3G (estrone-3-glucuronide) and PdG (pregnanediol glucuronide), which are hepatic metabolites of those hormones. The test strips infer the parent hormone's concentration from the metabolite level. Even blood immunoassays — the clinical gold standard for most hormone testing — do not detect hormone molecules directly. They measure the signal produced by labeled antibodies binding to the hormone: radioactivity in radioimmunoassays, fluorescence in enzyme immunoassays, chemiluminescence in others. The reported concentration is calculated from a standard curve. These assays are also subject to cross-reactivity — estradiol immunoassays can cross-react with over 100 metabolites, which is why LC-MS/MS has become the reference standard for accurate low-concentration quantification^[7].

The point is not that existing tests are unreliable — they are well-validated and clinically essential. The point is that virtually all hormone testing involves inference. The meaningful question is: how well-validated is the inference? How robust is it to confounders? What is the clinical utility of the output?

Clair's approach is a different modality of inference — physiological rather than biochemical — but the epistemic structure is the same: measuring one thing to determine another, validated against ground truth. Our validation shows 94.1% accuracy in cycle phase classification and 87% sensitivity for LH surge detection, benchmarked against clinical-grade hormone assays (urinary LH for surge detection, serum and urinary hormone panels for phase classification). As we advance through our clinical study program in the coming months, we will publish these validation metrics so the scientific community can evaluate our inference chain on its merits.

Can you actually detect estrogen and progesterone?

Each of these hormones produces a distinct, multi-system physiological signature — and our models are trained to recognize those signatures. But we want to be precise about what we mean and where the evidence is strongest.

Progesterone is the most physiologically "loud" and the easiest to detect. Its thermogenic effect on the hypothalamus, its influence on cardiac autonomic balance^[1][2], its impact on respiratory drive, and its role in extracellular fluid retention each leave measurable traces across different sensor modalities. When multiple signals shift together in the pattern specific to progesterone — temperature up, HRV down, heart rate up, bioimpedance shifted, sleep architecture altered^[16] — the inference is robust because no single confounder (exercise, alcohol, illness) produces that exact multi-system fingerprint. Our ablation studies show that removing temperature data alone degrades accuracy by approximately 16 percentage points — confirming that the model relies on multi-signal fusion, not any single channel.

Estrogen leaves a complementary but distinct cardiovascular signature. Rising estradiol during the follicular phase increases arterial compliance, shifts pulse wave morphology (detectable via multi-wavelength photoplethysmography at different tissue depths), and modulates autonomic tone toward parasympathetic dominance^[8]. These effects are well-characterized in the cardiovascular physiology literature and produce measurable changes in the spectral features of optical heart rate signals.

The critical innovation is simultaneous fusion of 130+ biomarkers across 10 sensor modalities, which allows the models to disentangle hormonal effects from the noise of daily life. This is why single-metric approaches — temperature alone, HRV alone — have never achieved what multi-modal sensing can.

Will Clair report actual hormone numbers (pg/mL, mIU/mL)?

Yes. Clair's sensor fusion foundational model estimates hormone levels from multi-modal physiological data. The underlying architecture is trained on paired wearable sensor data and clinical-grade hormone measurements, learning the mapping from multi-modal physiological embeddings to hormone concentrations. This capability exists today.

Clair 1.0 — our general wellness device — provides phase classification, hormonal trend tracking, and event detection (like LH surge and ovulation confirmation). Clair 2.0 will pursue FDA clearance and surface quantitative hormone estimates directly to users — estradiol in pg/mL, progesterone in ng/mL, LH and FSH in mIU/mL. The hardware is the same. Every Clair 1.0 device will receive the Clair 2.0 upgrade over the air, automatically and at no additional cost, once cleared.

The staging is regulatory, not technical. Reporting quantitative hormone values to consumers requires meeting FDA thresholds for accuracy and reliability, and our three-phase clinical study is designed to generate the large scale paired dataset needed to validate this at clinical grade. We will not surface numerical hormone reporting to users until it clears that bar.

What about confounders? Exercise, stress, caffeine, alcohol, illness — don't these affect the same signals?

Yes, they do. This is the single most important technical challenge in wearable-based hormone monitoring.

Heart rate variability is affected by psychological stress, sleep quality, exercise load, caffeine, nicotine, hydration, illness, and time of day. Skin temperature is affected by ambient temperature, alcohol, sleep environment, and physical activity. Electrodermal activity is affected by emotional arousal, physical exertion, and ambient humidity. Any single-metric approach to hormone inference would be overwhelmed by these confounders. This is why single-metric approaches do not work well for this application, and why we believe they never will.

Multi-modal sensing provides a fundamentally different approach to the confounder problem. The key insight is that confounders and hormones produce different multi-system signatures. Caffeine increases heart rate and may reduce HRV — but it does not shift bioimpedance, does not elevate temperature in the pattern progesterone does, and does not alter sleep architecture the way luteal-phase physiology does. Alcohol elevates temperature — but it does not produce the coordinated HRV, heart rate, and electrodermal changes that track with ovulation. Exercise increases heart rate and reduces HRV — but it does so acutely and with a motion signature that is trivially distinguishable from a resting hormonal shift.

When you measure across 130+ biomarkers spanning multiple physiological domains, the confounder and the hormonal signal are more separable in the combined feature space than in any single channel. Our models are being trained on data from real women living real lives — exercising, drinking coffee, sleeping poorly, getting sick — precisely so they encounter these confounders during training and learn to distinguish them from hormonal signal.

To be clear: our confounder robustness argument is supported by the real-world conditions of our training and validation data, and by the ablation studies showing multi-signal fusion outperforms any single channel. We have not yet published a formal confounder isolation study (systematically testing performance under controlled confounders like standardized caffeine doses or exercise protocols). Designing and executing such a study is part of our clinical program. In the meantime, the performance we report — including on irregular cycles where physiological stressors are more common — reflects real-world robustness, not laboratory conditions.

This does not mean confounders are irrelevant. Severe confounders — heavy alcohol use, acute illness, extreme exercise — can degrade performance. Our irregular cycle accuracy (84.3%) is currently lower than our regular cycle accuracy (95.2%), partly because irregular cycles often co-occur with physiological stressors that add noise. This is expected — irregular cycles are harder precisely because their hormonal patterns are less predictable. For context: BBT-based methods achieve approximately 72.5% fertile window prediction accuracy on irregular cycles with ovulation detection sensitivity dropping to 21%^[18], calendar and app-based methods achieve no better than ~21% accuracy for predicting ovulation day^[19], urine LH strips become unreliable for women with PCOS due to chronically elevated baseline LH^[20], and the most comparable published wearable ML system reported 79.85% fertile window prediction accuracy on irregular cycles with sensitivity of just 42.8%^[21]. Several commercial devices — including Ava — explicitly exclude irregular cycles entirely^[22]. Our 84.3% is not where we want it to be, but it reflects meaningful signal in a population where most existing methods struggle or do not attempt to operate. Improving performance on irregular cycles is a primary objective of our expanded clinical studies.

The Female Biology World Model

What is the Female Biology World Model?

A "world model" is a system that learns to understand and predict how a complex environment works — not by memorizing correlations, but by building an internal representation of underlying dynamics. A physics world model learns the laws governing how objects move. A weather world model learns atmospheric dynamics. The Female Biology World Model learns the dynamics of female reproductive physiology.

Specifically, it learns how the hypothalamic-pituitary-ovarian (HPO) axis operates as a coupled dynamical system: how GnRH pulses drive FSH secretion, how FSH stimulates follicular development and rising estradiol, how estradiol triggers the LH surge through positive feedback, how the LH surge triggers ovulation, how the corpus luteum produces progesterone, how progesterone inhibits GnRH through negative feedback, and how all of these hormonal events produce cascading physiological effects across thermoregulation, cardiovascular function, autonomic balance, body composition, sleep architecture, metabolism, and immune function — simultaneously and interdependently.

This is fundamentally different from a classification model that has learned "when temperature is high and HRV is low, it is probably the luteal phase." A classification model encodes correlations. Our approach learns dynamics — the temporal evolution of hormonal state and its multi-system physiological expression across hours, days, and cycles. We use the term "world model" because the system builds an internal representation of how the HPO axis operates, not because it has solved reproductive endocrinology. It enables the system to predict not just what phase you are in now, but how your physiology is likely to evolve, when hormonal transitions are approaching, and whether the pattern you are showing is consistent with normal cyclical variation or deviating from your personal baseline in potentially meaningful ways. The depth and accuracy of this internal representation will continue to improve as our training data grows — which is one reason our clinical study program is designed to be the largest paired wearable-and-hormone dataset in the field.

The model is trained on population-level data — learning the general dynamics of the HPO axis and its physiological expression — and then continuously adapts to each individual user's physiology. What is normal for one woman may be unusual for another. The model learns these individual differences through ongoing calibration, creating what is essentially a personalized digital representation of your hormonal biology.

How is this different from uploading my Oura data to ChatGPT?

This is a question we get often, and the answer reveals why building a domain-specific health AI is fundamentally different from applying a general-purpose language model to health data.

ChatGPT was trained on text. Our model was trained on physiology. When you upload a CSV of wearable data to ChatGPT, you are asking a system trained to predict the next word in a sentence to interpret continuous physiological signals. It has no built-in understanding of the HPO axis, no learned representation of how progesterone affects thermoregulation, no model of how estrogen modulates cardiovascular tone. It knows what the internet says about these topics — which is different from having learned the actual signal dynamics from paired wearable-and-hormone data.

Text tokenizers destroy physiological signal. Language model tokenizers are designed for words, not numbers. When a CSV row like "98.2°F, 72bpm, 45ms HRV" is tokenized, the numerical relationships and temporal dependencies are lost. The model cannot represent that 45ms HRV today versus 52ms HRV yesterday is a 13% decline that, in the context of a simultaneous 0.3°C temperature rise, signals progesterone dominance. It sees disconnected tokens.

There is no temporal dynamics model. Physiological state evolves continuously across multiple timescales — heart rate varies within seconds, HRV within minutes, temperature within hours, hormones across days. A world model learns these nested temporal dynamics and maintains an internal state that evolves over time. A language model processes each prompt as an isolated context window with no persistent physiological state representation.

There is no noise rejection. Wearable data is noisy. Temperature spikes from a hot shower. HRV drops from a movement artifact. Heart rate rises from climbing stairs. A domain-specific model trained on real-world wearable datasets has learned to distinguish physiological signal from sensor artifact and behavioral confounders. A general LLM has no mechanism for this.

There is no physiological constraint. A world model learns that certain states are physiologically impossible — progesterone cannot be high while temperature is low in a healthy cycle, FSH and estradiol have specific temporal relationships. These constraints emerge from training on real physiology. A language model can hallucinate physiologically impossible predictions because it has no grounding in biological reality.

Domain-specific health foundation models consistently outperform general-purpose LLMs on health prediction tasks, precisely because they learn the actual dynamics of the systems they model rather than pattern-matching on text descriptions^[23].

How Is This Different?

If Oura and Whoop measure heart rate, HRV, and temperature too, why can't they do this?

The difference is more fundamental than hardware specs. It is a difference in what the system was built to understand.

The paradigm is different. Oura and Whoop are fitness and recovery platforms. They were designed to answer questions like: "How well did I sleep?" "Am I recovered enough to train?" "What is my resting heart rate trend?" Their algorithms were built to find trends in data — and when the underlying trend is itself a cycle, as it is in female physiology, they treat cyclical variation as noise to be smoothed over. A woman's HRV dropping 12% in her luteal phase looks like a recovery problem to a fitness algorithm. Her resting heart rate rising 3.5 BPM after ovulation looks like declining fitness. Her sleep efficiency changing in the premenstrual window looks like a sleep hygiene issue. It is none of these things. It is normal hormonal physiology — and a system that does not understand this will systematically misinterpret half the population's data.

Clair was built from the ground up to treat female hormonal physiology as the primary signal, not background noise. Our models understand that cyclical variation is the core biological reality — that the body's shifting cardiovascular, thermoregulatory, autonomic, metabolic, and sleep patterns across the menstrual cycle are not deviations from a baseline but expressions of the underlying hormonal state. This is not a feature layered onto a fitness tracker. It is the foundational design principle.

The sensing is different. This paradigm demands a different sensor stack. Consumer wearables optimize for a handful of metrics — multiplexed PPG for heart rate, basic thermistors for temperature trends, accelerometers for motion. These are sufficient for fitness tracking but insufficient for resolving the subtle, multi-system physiological signatures that differentiate hormonal states from everyday confounders. Clair includes proprietary sensing modalities that consumer wearables do not have. From these sensor modalities, we extract over 130 mechanistic biomarkers. This breadth is what enables the system to distinguish hormonal shifts from confounders — because confounders do not produce the same coordinated pattern across all these channels.

The intelligence is different. Even with the right sensors, you need a model that has learned what hormones actually do to the body — not from text on the internet, but from paired sensor-and-hormone data. Our models are trained on continuous multi-modal wearable data aligned with clinical-grade hormone measurements. They learn the actual mapping between physiological signatures and hormonal states. General fitness algorithms do not see hormonal patterns because they were never trained to look for them, their training data never included hormone ground truth, and their architecture was not designed to model the coupled multi-system dynamics of reproductive endocrinology.

How is this different from just tracking basal body temperature?

BBT tracking captures a single consequence of a single hormone (progesterone's thermogenic effect) with a single measurement per day^[3]. It confirms that ovulation likely already occurred — retrospectively. It cannot tell you which phase you are approaching, when ovulation is imminent, or whether ovulation actually happened versus an anovulatory temperature fluctuation.

Clair captures the full multi-system physiological response to the entire hormonal cycle — across multiple sensing modalities, 130+ biomarkers, continuously — and uses models trained against hormone ground truth to classify phase, detect hormonal events, and (in Clair 2.0) estimate hormone concentrations. The difference is not incremental. It is architectural.

There is also a robustness advantage. Temperature is easily confounded: alcohol, illness, poor sleep, ambient temperature, and exercise all shift it independently of hormones. Multi-signal fusion provides natural noise rejection because confounders rarely produce the same multi-system pattern as hormonal changes. When temperature is elevated but HRV, cardiac autonomic balance, bioimpedance, and electrodermal activity are unchanged, the system can identify the elevation as non-hormonal.

Clinical Validation

What clinical evidence do you have, and how should we evaluate it?

Our prototype was validated on 40+ women across 127 complete menstrual cycles, generating over 5,000 days of continuous physiological data. Wearable predictions were validated against clinical-grade hormone assays — not self-reported period dates, not calendar-based assumptions.

This distinction matters. A common weakness in menstrual cycle tracking studies is reliance on self-reported menses as ground truth, which research has shown can be wrong in up to 50% of phase assignments^[9]. Our validation used actual hormone measurements as the reference standard.

Key results:

• 94.1% overall accuracy in classifying cycle phase from wearable sensor data alone — no blood draws, no urine tests, no user input required. This figure was obtained using leave-one-subject-out cross-validation — the most conservative evaluation scheme, where the model is tested on individuals it has never seen during training (not just held-out cycles from known individuals). Leave-one-cycle-out cross-validation (testing on new cycles from known individuals, a less conservative scheme) produced higher accuracy. We report the leave-one-subject-out figure because it better reflects real-world performance on new users.
• Phase-specific performance: Menstrual (96% sensitivity, 98% specificity), Follicular (92% sensitivity, 95% specificity), Ovulatory (93% sensitivity, 97% specificity), Luteal (94% sensitivity, 96% specificity). For context, in a 4-class problem with unequal phase lengths (the luteal phase alone spans roughly 40–50% of the cycle), a naive classifier that always guesses the most common phase would achieve approximately 40–50% accuracy depending on cycle length. Our 94.1% represents roughly double that baseline — substantial signal above chance.
• 87% LH surge detection sensitivity (93% specificity) with timing accuracy within approximately one day.
• The study included both regular (71%) and irregular (29%) cycles, ages 18–45, BMI range 18.5–35, and diverse skin tones (Monk Skin Tone shades I–X).
• We also built models to track progesterone patterns to confirm whether ovulation actually occurred — critical because anovulatory cycles are more common than most people realize, especially during stress or with age.

Where the evidence is weaker — and we are honest about it. Performance on irregular cycles (84.3%) was meaningfully lower than regular cycles (95.2%). This is expected — irregular cycles are harder precisely because their hormonal patterns are less predictable — but it means women with PCOS or other cycle irregularities will experience reduced accuracy with the current model. Improving performance on irregular cycles is a primary objective of our expanded clinical studies. We are fully aware that larger, more diverse populations are needed — which is exactly what our clinical study program is designed to provide.

Aren't you just detecting circadian patterns that correlate with cycle phase, not actual hormonal effects?

This is a legitimate scientific question, and it deserves a direct answer.

Hormonal rhythms and circadian rhythms interact bidirectionally — the circadian clock influences hormone release, and hormones influence circadian function^[10]. A model that simply learned "what time of cycle is it?" from circadian signatures would produce false predictions that track calendar time, not actual hormonal state.

Two aspects of our validation address this concern. First, our models are trained against actual hormone measurements as ground truth — not calendar day, not days-since-last-period. The model learns the mapping from physiological signals to hormonal state as measured by clinical assays. If it were merely learning calendar patterns, it would fail when the actual hormonal timeline diverges from the expected timeline — which is exactly what happens in irregular cycles. The fact that the model maintains 84.3% accuracy on irregular cycles (where calendar-based prediction fails entirely) is evidence that it has learned genuine hormonal signal, not just cycle-day correlations.

Second, our leave-one-subject-out cross-validation tests the model on individuals it has never seen. If the model were memorizing subject-specific circadian patterns, it would fail on new individuals. The generalization across subjects suggests it has learned population-level hormonal physiology, not individual behavioral schedules.

Is this proof that confounding with circadian patterns is zero? No. Disentangling circadian and hormonal contributions to physiological signals remains an active area of research. But the evidence supports that our models are capturing genuine hormonal signal — and the clinical study program is designed to further characterize the relative contributions of each.

What does your clinical study roadmap look like?

We are running a three-phase clinical study program designed to both solve the data problem in non-invasive hormone research and rigorously validate our system under independent scrutiny.

Phase 1 — Data Collection & Model Training (enrollment begins April 2026)

This study pairs continuous Clair wearable data with clinical-grade hormone ground truth collected at multiple time points throughout each day. The goal is to assemble the largest and most granular paired wearable-and-hormone dataset ever collected for menstrual cycle research — the kind of dataset this field has needed but has not had. Existing public datasets in this space are either small (fewer than 50 participants), use low-frequency hormone sampling, or lack the multi-modal sensor depth necessary for robust model training. This study is designed to close that gap.

Phase 2 — Model Refinement & Robustness Validation

Using the paired dataset from Phase 1, we retrain our models on production hardware, validate performance across diverse cycle types (regular, irregular, short, long), and stress-test real-world robustness. This phase will also expand the participant population to include groups not represented in our initial study — including women in perimenopause and women on hormonal therapies.

Phase 3 — Independent Validation at Stanford (BeeHive Initiative)

The final phase is a fully independent validation study on 150 women through the Stanford Gladstone BeeHive Initiative. This is not us evaluating our own system. It is independent researchers, under rigorous clinical and institutional review board oversight, validating Clair's performance against hormone ground truth — with results submitted for peer-reviewed publication. We pursued this path because claims about understanding the human body should meet a high evidentiary bar, and that means letting others check your work.

Our goal extends beyond product validation. We want to advance the science of non-invasive hormone monitoring for the entire field.

Is Clair FDA approved?

Clair 1.0 will launch as a general wellness device, which does not require FDA clearance under current regulations. It provides cycle phase tracking, hormonal trend insights, and wellness metrics.

Clair 2.0 will pursue FDA 510(k) clearance to enable medical-grade claims — including quantitative hormone reporting and ovulation detection for fertility applications. The regulatory pathway is well-established within the existing classification framework for reproductive health monitoring devices.

Both Clair 1.0 and 2.0 share the same hardware and the Clair 1.0 users will be automatically upgraded to 2.0 free of cost via OTA firmware update.

Perimenopause, Menopause & Hormonal Therapies

Does Clair work for women in perimenopause or menopause?

The current system is designed around the physiology of the HPO axis: the cyclical interplay of GnRH, FSH, LH, estradiol, and progesterone, and the downstream physiological effects each of these hormones produces.

During perimenopause, the HPO axis does not shut down — it becomes dysregulated. Follicular reserves decline, inhibin B drops, FSH becomes elevated and erratic, estrogen fluctuates between undetectable and supraphysiological levels, and an increasing proportion of cycles are anovulatory. These are not invisible changes. They produce dramatic physiological effects: thermoregulatory instability (hot flashes produce distinct skin temperature, electrodermal activity, and heart rate signatures that published research has shown can be detected — and even predicted — by wearable sensors^[13][14]), significant declines in heart rate variability^[15], and disruptions in sleep architecture, circadian temperature rhythms, and autonomic balance.

The physiological signals are there — arguably even louder than during regular cycling. Our Phase 2 and Phase 3 clinical studies will include these populations more prominently to test how the system performs for that population.

What about women on hormonal contraceptives or hormone replacement therapy?

Women on hormonal contraceptives and HRT were excluded from our initial validation study, because the physiological effects of exogenous hormones are distinct from endogenous cycling, and we needed to validate the foundational approach first.

There is strong evidence that exogenous hormones produce detectable physiological signatures. Women on hormonal contraceptives show consistent temperature elevation similar to the luteal phase, blunted nocturnal temperature drops^[11], and measurable changes in HRV and sleep architecture^[16]. Research on HRT shows that exogenous estradiol affects thermoregulation (promoting vasodilation and lowering core temperature), while progestin components modify cardiac autonomic balance in the opposite direction^[17]. Different HRT formulations — transdermal estradiol versus oral, micronized progesterone versus synthetic progestins — produce measurably different physiological profiles.

These are physiological effects that our sensor stack is well-positioned to detect. The open question is not whether the signals exist — it is whether our models can learn to interpret them accurately in these populations. That requires training data with appropriate hormone ground truth, which our subsequent study phases will collect.

We take this seriously because a large proportion of women who would benefit from continuous hormonal monitoring are on some form of hormonal therapy — and a system that excludes them is incomplete.

General Capabilities

Can Clair track workouts, cardiovascular health, sleep, and other wellness metrics?

Yes. Clair's sensor array is more comprehensive than most consumer fitness wearables, not less. We continuously capture heart rate, heart rate variability, skin temperature, multi-axis motion, sleep staging, respiratory patterns, electrodermal activity, bioimpedance, and more.

You will get workout detection, cardiovascular health metrics, sleep analysis, and stress monitoring. But with Clair, these metrics are contextualized by your hormonal state.

This matters because hormones are the single largest source of unexplained variability in the metrics other wearables report. Your HRV drops in your luteal phase — not because you are stressed, but because progesterone shifts autonomic balance^[1][2]. Your resting heart rate rises by 3–4 BPM after ovulation — not because your fitness declined, but because your physiology changed. Your sleep quality decreases in the late luteal phase — not because of poor sleep hygiene, but because progesterone metabolites affect GABA receptor activity and alter sleep architecture^[16].

Without hormonal context, these shifts look like noise — or worse, like problems that do not actually exist. Clair gives you the full picture.

Intellectual Property & Transparency

Why haven't you published the full technical details and peer-reviewed studies?

We have shared more of our approach publicly than most companies at our stage — this post, our blog, and our validation results. We are also running a clinical study program specifically designed to produce independent, peer-reviewed publications in the coming months and years.

But there are specific technical innovations — in our sensor architecture, our signal processing pipeline, and our machine learning approach — that are currently in the patent process. We have a responsibility to protect these innovations before making those details public.

As our patents are secured, we will be able to share more. In the meantime, the independent validation study at Stanford through the BeeHive Initiative is designed so that the science can be evaluated on its merits by researchers outside our organization — with full methodological transparency in the published results.

Clair is building the first non-invasive continuous hormone monitor. Learn more at wearclair.com.