Make the invisible
accountable.

01 / OBSERVE

What can we actually measure?

This prototype describes the appearance and the image motion of a region. It calculates relative brightness, colour variation, edge structure, the dominant axis of that structure and how coherent it is, frame-to-frame image change, and, where the image supports it, displacement: how the region moved relative to the whole view. The same calculations can describe a plant, a face, a moving object or an environment. These statistics do not classify entities or explain their internal state. Optional on-device models separately estimate class labels, masks and boxes; labels can be wrong and never identify an individual.

Captured pixels → spatial descriptors and fields → model-estimated displacement → timestamped Aura state → visual grammar → evidence. Synthetic fixtures and camera inputs share this entire pipeline.

A multidimensional state, with provenance

A(t) = {region, entity, fields[adaptive grid], flow[adaptive grid], global{residualMotion, motionAgreement, wholeFieldMotion, textureAlignment, activityConcentration, appearancePersistence, chroma…}, evidence, cameraMotion, temporal, quality, time}

Every channel carries its value or an explicit unavailable state, its method, unit, evidence class, timestamp, window, region and confounds. Missing evidence is null, never zero. A silhouette, when present, is a model estimate with model id, hash and age; without it the state says so.

02 / REVEAL

Eight ways to make evidence visible.

These are competing experiments, not a selected identity. Five WebGL2 grammars read the measured fields, tangents, luminance and displacement; three earlier Canvas grammars remain as controls. Switch while a study runs, pause on a state, or restart an identical synthetic sequence for comparison.

Signature invariants of the WebGL2 grammars

1. Around, not on. The aura is a layer outside each subject’s silhouette: a membrane hugging the contour, an atmosphere of filaments wrapping it, falling off with distance and reaching further when the subject moves. The soft band can overlap the estimated edge. Geometry comes from a signed distance field followed by smoothing, which changes exact distances. Segmentation masks are model estimates; rounded detector boxes and moving regions are graphical proxies for shape; its extent and brightness from measured motion and structure. Between detections a subject’s box is carried by raw image displacement measured inside it and snaps back on the next detection; motion during inference is not yet reconstructed; the sheet says when this is happening.
2. Polarity, defined. The rim’s colour, width and continuity encode one number: the coherence of a subject’s observable dynamics over the last few seconds. It is a weighted mean of how much its appearance persists, how evenly paced its movement is, how smooth that movement is in the frequency domain, whether the movement goes in one direction, how legible its structure is, and how good the acquisition was. Movement that holds together reads positive (warm, continuous, wider); movement that arrives in bursts, fragments and disagrees with its own direction reads negative (cool, broken, tighter). Quantity of movement is measured and shown but deliberately kept out of the number, because high activity accompanies opposite states. Missing evidence removes the rim rather than colouring it neutral. It is not emotion, mood, health, identity, intent, character or worth. How it is computed, and what it cannot be.
3. Unsigned unless measured. Marks follow the axial tangent of image structure; a direction appears only where the displacement estimate passed its gates.
4. Evidence-proportional ink. Edge strength, orientation coherence, acquisition quality and freshness multiply the ink. Null channels remove their contribution.
5. Family colour distinguishes regions. Outside any subject, hue is a fixed function of a mark’s orientation (20° + 140°·|sin θ|). Each detected subject is drawn in its own hue family from a designed palette, stable while that subject stays in view and distinct from its neighbours, so several auras can be told apart; the same family appears as a chip beside the model label. Chroma follows mean Oklab chroma inside the estimated subject box. The separate rim colour encodes the dynamics index described above. Neither family nor rim colour establishes mood, health or worth.
6. Luminance-only, lighter than the background. The layer is composited as added light (screen blending), so it cannot darken the world; over a very bright surface it fades instead. Legibility on an additive glasses display remains untested. Aperture uses normal blending for its camera refraction.
7. Nothing moves on the clock. The integral texture advances only when a new sample arrives; between samples the display interpolates that sample’s measured displacement and then stops. Staleness fades the field and says so.

Base field extent, noise scale, sheet spacing and refraction cap are design constants. Subject residual motion adjusts the shell radius, and local edge density adjusts filament spacing. No depth, energy, emitted luminosity, physical reach or heartbeat is measured.

Open the experience → · Return to the studies ↗

02B / READ

Positive and negative, defined.

“Positive aura” and “negative aura” are ordinary words for something people really do: forming a fast impression of how someone or something is, from a few seconds of watching. That impression is not nothing — a 1992 meta-analysis of thirty-eight studies found judgements from behaviour samples under five minutes predicted real outcomes, and were barely improved by watching for longer. Aura takes that seriously and then refuses the next step. What a pattern of movement means is not fixed: a 2019 review of the evidence on reading emotion from facial movement concludes that a given configuration does not reliably signal a particular emotional state across people and situations, and recommends describing the movement rather than naming a feeling. A survey of bodily expression reaches a similar position for the body.

So Aura describes the observable dynamics of a region of the image, with units, and never names a feeling. Everything below is a property of pixels over time.

What the number is

Over the last few seconds of evidence, Aura measures six things inside a subject’s region and combines them into one index between −1 and +1.

The index is 2 × (weighted mean − 0.5). Below three available dimensions, or with poor acquisition, it is unavailable and the rim loses its colour entirely rather than settling at neutral.

What is measured but deliberately left out

How much something moves. A 1998 study of 224 acted portrayals found high movement activity, expansive movement and high movement dynamics in both elated joy and hot anger, while sadness showed low activity and a collapsed posture. Quantity of movement therefore cannot carry a sign. Aura shows activity as its own number, next to the index, never inside it.

Apparent size. Expansiveness is a real descriptor, but on a single camera it is inseparable from simply coming closer. Shown, with the confound stated; excluded from the index.

Rhythm. When the movement repeats with a period between half a second and eight seconds, Aura reports the period and how much of the band it dominates. A fan, a gait, a pendulum and a handheld camera all produce one. It is not breathing, not a pulse, and no health quantity is derived from it.

Between two subjects

With two subjects in view, Aura reports the peak correlation of their two motion series within one and a half seconds, and which one leads. Coordinated movement between people is measured this way in psychotherapy research, where it has been associated with relationship quality. Aura reports the correlation and stops there: it is not rapport, liking, agreement or a relationship.

What it is not

• Not calibrated. The weights and thresholds are engineering choices. There is no ground truth for “coherence of dynamics”.
• Not a person. The measured region is a model’s box or mask; it usually includes background and can lag a moving subject.
• Not a physical field. Kirlian photographs, the usual visual reference for an aura, record a corona discharge that tracks moisture, contact pressure and voltage. Aura records no emission.
• Not tested with people yet. Whether the rim communicates coherence, or quietly invites “this person is angry”, is an open question and the next gate.

That last point is not rhetorical. A four-minute study in this prototype asks whether anyone can tell two layers apart, whether they notice when there is no reading at all, and whether the edge quietly reads as a feeling. A device self-test measures what this handset can actually run.

Barrett et al., inferring emotion from facial movement · Kleinsmith and Bianchi-Berthouze, affective body expression · Wallbott, bodily expression of emotion · Balasubramanian et al., SPARC · Tremoulet and Feldman, animacy from motion · Kirlian photography

02C / A PERSON, RIGHT NOW

Expressed mood, from the research.

Point Aura at a person and it estimates the mood their face, posture and movement express right now, on the model psychology uses most: Russell’s circumplex, where every mood sits on two axes — valence (unpleasant ↔ pleasant) and arousal (calm ↔ activated). The answer is a word from the circumplex — content, excited, alert, tense, uneasy, subdued, quiet, relaxed or neutral — or one of three configurations the literature names directly: joyful, surprised, focused. The mood map shows where the person sits and the path they took over the last minute; under it are five signals: positivity, energy, engagement, visible tension and openness.

Thin evidence stays near neutral, no mood word appears without a visible expression, posture or movement cue (facing the camera alone is not a mood), and a new word must hold for about a second before it replaces the last one. Every reading shows its cues, each with its source, and an evidence level. The weights are a transparent first design; none has yet been checked against people’s own reports of how they feel.

This is the mood a person’s behaviour expresses — how they come across — not what they feel. The same face comes from different feelings across people and situations (Barrett, Adolphs, Marsella, Martinez & Pollak 2019).

Each face, and each pair

Faces differ at rest — some mouths turn down, some brows sit low — so after about 20 seconds Aura learns each person’s resting face and reads movement away from it, the person-specific normalisation used in automatic facial analysis (Baltrušaitis, Mahmoud & Robinson 2015). A held smile or frown is never mistaken for a resting face. Cues count for less when the face is turned away or the light is poor, and the reading says so. When two people’s smiles or movements rise and fall together, a bridge joins them: “in sync”. Synchrony of this kind goes with rapport in studies (Ramseyer & Tschacher 2011; Hove & Risen 2009; Hess & Fischer 2013); Aura measures only the coordination, checks it against the same signals shifted out of alignment, and never says how the two feel about each other.

The aura follows the mood

The layer around a person takes the colour of their place on the mood map — gold and copper for pleasant, rose for excited, teal for alert, violet for tense and uneasy, green for relaxed and quiet; neutral keeps the subject’s own colour — and its rim shows their mood aura, warm and continuous when positive, cool and broken when negative. A ring around the face works as a compass: its marker sits where the person is on the mood map. On the person being read, the features behind the reading glow and callouts name each cue with its source; when a cue appears, Aura says so. These colours are a designed encoding of the reading, not a property of the person.

What it will never say

• Anything about personality, character, intelligence or trustworthiness. Impressions of traits from a face are consistent between observers but not accurate, and systems that claim otherwise reproduce bias.
• Anything about health, stress or mental state as a condition.
• Who someone is. No face is recognised, and nothing that could identify anyone is computed or kept.
• A reading without evidence. With too little face or posture in view it says “Reading…”, and it never calls unmeasured movement stillness.

Everything runs on the device, and the camera stops whenever the app is not in front of you. People being read may not know it: point it at others with their awareness.

No buttons

Aura opens with a short welcome while the camera and the models start — each step lights up only when it has really started — and then opens into the live view by itself; a tap skips it, nothing needs one. It finds people and things by itself, follows the most relevant one — a person first, then a named object, then whatever stands out — and keeps its reading on screen. A face or body counts as a person only once it is found consistently, so a patterned object is not read as someone. You can also tap anything — it is optional, never required: Aura locks on to what is under your finger (or looks for what stands apart right there), and its reading appears in a card beside it: the aura verdict on a needle from negative to positive, the mood or pattern, what it rests on, and three measurements. Tap it again, tap the ×, or tap empty space to let go. The first time, a ghost hand shows how; the menu at the top left holds the settings (effects, labels, sync, glow, mirror, glasses layout), a live view of each step of the detection, and the research. On glasses (a 600 × 600 display) the camera image is not drawn, so the aura and a larger reading float over the world. Whether a glasses display app can receive camera frames, and aligning the aura with what the wearer sees, are not established yet.

Open Aura →

03 / UNDERSTAND

Knowing when we cannot know.

Quality flags are engineering checks, not calibrated confidence probabilities. Missing data stays unavailable. A dim field in low light means less usable image evidence; it says nothing about the strength or vitality of the subject.

• Camera motion, exposure, autofocus and lighting can produce apparent activity.
• A selected region does not establish an object’s boundary. Automatic masks and boxes carry model labels and age; the optional tap-seeded silhouette carries its model and age. Re-association and box carrying can drift and do not establish identity.
• Whole-field motion cannot distinguish a moving camera from a region that moves as a whole. Gyroscope rotation adds context, but low rotation cannot rule out camera translation; the ambiguity remains.
• RGB does not measure depth, temperature, thoughts, intent, morality, honesty, health or compatibility.
• Persistent image descriptors do not establish calmness. Facial movement does not establish felt emotion. Review of facial-expression inference.
• rPPG is a separate research track. Real-world measurement variability, movement and illumination prevent us from treating a camera-derived rhythm as a dependable physiological measure here. No heart rate or health result is shown. Real-world rPPG evaluation.
• Visual beauty and reproducible calculations do not validate a new scientific construct. User comprehension, cross-device reliability and real-world usefulness remain untested.

04 / OBSERVE WITHOUT INTRUDING

The device is the boundary.

Camera access begins only when you select Open camera or Open Aura and your browser permits it. RGB frames are analyzed in memory on your device. The application does not upload or record camera frames, persist observations, identify people, or request microphone or location access. It uses no analytics or external font services.

Three optional on-device models (MediaPipe Tasks Vision 1.0.1, Apache-2.0: DeepLab-V3 for automatic subject regions of 20 named classes, EfficientDet-Lite0 for boxes of 80 everyday classes, and the Interactive Segmenter for whatever you tap) are served from this site and verified by hash before use; they run in Web Workers on your device. Their labels are model labels, shown as such; regions are re-associated frame to frame by overlap, which is not identification. The workers refuse any network request that does not go to this site, and report every attempt, so the runtime's telemetry cannot leave your device whatever the host's headers say. Its runtime attempts to send usage telemetry to Google; this site’s Content-Security-Policy blocks that request, and the worker reports any such attempt. Anyone hosting this code must send the same policy.

Stop camera releases every media track and clears the current evidence. Hiding or leaving the page stops capture. Returning requires an explicit restart. Pausing a live camera stops it; it does not leave a hidden stream running. Sensor failures never silently switch to synthetic data.

The hosting service still delivers files and may process ordinary access logs and authentication. “On device” describes the sensing pipeline, not the absence of website requests. Observe people and private spaces with their awareness and permission. Aura never requires touching, disturbing or actively illuminating a living being.

05 / BEYOND THE SCREEN

Built for the browser. Open to the spatial world.

Today: a browser with JavaScript and Canvas 2D runs the synthetic studies. Camera mode additionally requires HTTPS (localhost is allowed), MediaDevices camera support and permission. The prototype feature-detects video-frame callbacks and has a current-time fallback. Device compatibility is a test matrix, not a blanket guarantee.

Meta Display's current web guidance describes a 600 × 600 additive display: black is transparent. This supports sparse geometry but does not itself establish world-anchored entity overlays. Camera access, localization and display anchoring must be verified separately. Meta display development, developer availability, display guidelines, Quest camera API.

06 / EARN THE NEXT CLAIM

Research gates, not feature promises.

1. Optical baseline. Verify deterministic measurements, missing states, timing gaps and source changes. This is the implemented starting point.
2. Visual comparison. Replay the same traces. Vary one channel at a time, freeze input, and compare against a simple static baseline. Reject visual complexity that adds no comprehension.
3. Device and confound study. Test camera pan, subject motion, flicker, darkness, blur, multiple entities and occlusion across real devices. Measure age of evidence, dropped samples and thermal behavior.
4. Comprehension study. Counterbalance experiment order. Ask users to identify changing observables and unavailable data. Track misleading interpretations separately from aesthetic preference.
5. Entity perception. Add evaluated detection/segmentation and ephemeral association. Validate each supported entity category; abstain elsewhere.
6. Spatial and glasses research. Add coordinate transforms, occlusion and anchoring only on capable hardware. Validate additive-display legibility and registration drift.

No human-subject study or hardware evaluation is claimed. The initial proposed comprehension gates are at least 80% pairwise recognition of observable differences after a legend and 90% recognition of unavailable evidence. These are proposed acceptance criteria, not achieved results.

Sources were checked on 12 September 2026; the mood-model sources on 19 September 2026. Technical definitions support the measurements; they do not validate Aura's visual mappings or inferential claims. Displacement estimates were verified on synthetic fixtures with known ground truth; no phone or participant study has been run.

07 / PRACTICAL QUESTIONS

A few things to know.