Research foundations & methods

A transparent, detailed account of the science InnerLight is built on, the technologies it uses, and why — written for researchers, clinicians, and reviewers. InnerLight itself is not yet validated in a controlled trial; this page documents the established principles behind its design and our commitment to testing it honestly.

1. Calming sound — the Iso-Principle

InnerLight’s use of sound is built on the Iso-Principle from music therapy: meet a person’s current emotional state with matching music, then gradually shift the music toward calm to carry them with it. This is a long-standing clinical method with controlled experimental support.

Starcke K., Mayr J., von Georgi R. (2021). “Emotion modulation through music after sadness induction — the Iso principle in a controlled experimental study.” International Journal of Environmental Research and Public Health, 18(23).

Music with auditory beat stimulation RCT protocol (2025). BMJ Open, 15(6):e094784 — describes Iso-principle personalization against baseline Self-Assessment Manikin (SAM) scores.

2. Target tempo for relaxation (60–80 BPM)

Research indicates that music in the 60–80 beats-per-minute range supports relaxation by aligning neural oscillations (alpha-wave activity) with the musical rhythm, shifting arousal from tense toward calm. InnerLight prioritizes tracks and, in development, dynamic tempo shaping toward this range.

Xu R., Li J. (2025). “AI-driven music intervention based on five-tone theory for anxiety: a preliminary pre-post feasibility study.” Frontiers in Psychology, 16:1669029. (Real-time HRV-guided tempo modulation.)

Frontiers in Digital Health (2025), 7:1552396 — review of music therapy, entrainment, and AI-driven biofeedback.

3. Real-time, physiology-guided adaptation (in development)

The strongest current evidence favors adjusting musical parameters — tempo, volume, complexity — smoothly and in real time in response to physiological signals, rather than abruptly switching tracks. When tension rises, effective systems slow the tempo and simplify the music with soft transitions. This is the direction of InnerLight’s ongoing sound development. We are deliberately re-examining which signal should drive it: because beats-per-minute is an unreliable indicator of emotional state, we do not use heart rate as a distress signal, and are building a more trustworthy read of how a person is doing.

REMAST: Real-time Emotion-based Music Arrangement with Soft Transition (arXiv:2305.08029).

Williams et al. (2020); Jiao (2025) — adaptive functional music generation with real-time biofeedback, reviewed in Frontiers in Psychology (2026), 16:1741463.

4. Contactless heart reading — remote photoplethysmography (rPPG)

InnerLight reads heart rate from a standard webcam using remote photoplethysmography: detecting the tiny color changes in facial skin as blood pulses beneath it. We combine forehead and cheek skin regions (avoiding the eyes and mouth, which introduce motion noise), verify skin pixels, detect the beat period by autocorrelation, and apply physiology-informed smoothing so implausible jumps are rejected. In low light, the signal is automatically brightened (adaptive gamma correction) before analysis so people in dim conditions are not excluded.

Method basis: chrominance- and plane-based rPPG (POS/CHROM family); forehead and cheek regions of interest shown to carry strong pulsatile signal in systematic reviews of rPPG ROI selection.

Low-light handling follows gamma-correction and histogram-based enhancement approaches evaluated for rPPG under poor illumination.

Why webcam rPPG, and not a wearable or a specific product: a crisis tool must work for anyone, instantly, with no device to buy, pair, or install. Wearables and clinical pulse oximeters are more accurate but exclude anyone who doesn’t own one in the moment. Deep-learning rPPG models are strong but require a server and heavy computation. Browser-based rPPG is the only approach that runs immediately for everyone on a phone or computer — so we use it, and we are transparent about its limits: it needs reasonable light and a mostly still face, and we label every reading by confidence (measured / estimated / baseline-held) rather than overstating precision.

5. Facial-signal reading — MediaPipe

For facial-expression signals InnerLight uses Google’s MediaPipe Face Landmarker, which measures dozens of specific facial-movement values (blendshapes) rather than guessing a single emotion label. We chose MediaPipe because it is free, runs entirely in the browser (no images ever leave the person’s device for this), is well-documented, and is widely used and maintained — important for a tool that must be reproducible by a research team.

6. Grounding through real imagery

InnerLight uses real photographs, not animation, as grounding scenes. Realism is used deliberately: concrete sensory grounding is a recognized technique for interrupting distress and dissociation and returning attention to the present.

7. Privacy & encryption

Privacy is foundational, not an afterthought. Specifically:

• Session content is not stored in raw form. Only a summary a person chooses to save is retained, and identifying details (emails, phone numbers, handles, long digit strings) are automatically removed before any research record is stored.
• Returning-user memory is encrypted with a key derived from the person’s own return code. Their saved story cannot be read without that code — not by us, not by anyone with server access. If the code is lost, the data is unrecoverable by design. This uses the Axiom Harmony Protocol, our applied encryption layer, so that the person — not the operator — holds the key to their own story.
• The live biometric monitor is anonymous and ephemeral. It shows heart rate and calm-state under anonymous labels (Person 1, Person 2), holds no words, and expires shortly after a session ends.
• Facial analysis runs on-device. The person’s video is analyzed in their own browser for heart and expression signals; the raw video is not transmitted for that analysis.

7b. The Axiom Harmony Protocol (AHP) — what it is, and how it protects you

The Axiom Harmony Protocol (AHP) is InnerLight’s encryption layer — the system that turns anything a person chooses to save into scrambled, unreadable data that only they can unlock. It is not a marketing name over weak protection; it is built on the same class of encryption used to protect banking and government data. Here is exactly how it works, first in plain terms and then in technical terms.

In plain terms: when you save your story, InnerLight takes your private return code and, through a deliberately slow mathematical process, turns it into a unique digital key. It then locks your words with that key so thoroughly that the stored result looks like random noise. Your code is the only thing that can produce that key again. We never keep your code, so we can never unlock your story — and neither can anyone who breaks into the server. If you lose the code, the data is gone for good. That is the trade-off of true privacy: the lock is real, and you hold the only key.

In technical terms, for reviewers: AHP encrypts each payload with AES-256-GCM — the Advanced Encryption Standard at 256 bits in Galois/Counter Mode, an authenticated cipher that protects both confidentiality (no one can read it) and integrity (tampering is detectable). The key is derived from the person’s return code using PBKDF2-HMAC-SHA256 with 390,000 iterations and a random salt — a deliberately slow key-stretching function that makes brute-force guessing enormously expensive. Every encryption uses a fresh random nonce, and the protocol version is bound in as authenticated associated data. The key is never written to disk; only the ciphertext, salt, and nonce are stored. This is a zero-knowledge design toward the operator: InnerLight holds encrypted bytes it cannot read.

Honest limit: AES-256-GCM with strong key derivation is robust modern cryptography, but it is not yet post-quantum. A documented future hardening path is to add a post-quantum key-exchange layer (for example, ML-KEM / Kyber) alongside the authenticated symmetric encryption. We state this openly rather than overstate the protection.

8. What we measure, and how we stay honest

InnerLight records anonymous, aggregate research metrics designed around recognized digital-health frameworks: uptake, engagement, session duration, adherence, and completion, alongside expression shifts, sound responses, self-reported calm (a wordless Self-Assessment Manikin scale), and heart-rate trends measured against each person’s own baseline. Every heart reading carries a confidence tier so coverage is complete without overstating precision. We follow the scientific method explicitly: a falsifiable hypothesis, stated predictions, an instrument that gathers the data, and a commitment to replication and peer review.

9. The mission’s evidence — alternative crisis response

InnerLight’s north star is to become a trusted first call for a person in crisis — a path to real help that does not begin with an armed response. This is not a solitary idea; it is a documented national movement with a growing evidence base. InnerLight is a civilian, software front door to that movement: it bridges people toward care-first human responders, and toward 911 only when immediate physical danger requires it.

Dee T.S., Pyne J. (2022). “A community response approach to mental health and substance abuse crises reduced crime.” Science Advances, 8(23). (Denver STAR: 34% reduction in low-level offenses in served neighborhoods.)

RTI International (2026). Quasi-experimental evaluations of Durham’s HEART and Greensboro’s BHRT programs: unarmed response teams matched police response times, reduced arrests, and increased connections to supportive services.

Human Rights Watch (2026). “Self-Determination is the Pathway to Liberation” — national assessment of non-police, consent-based, rights-respecting crisis response programs.

Vera Institute of Justice (2022). Civilian Crisis Response Toolkit — documents community distrust of police-linked emergency lines and recommends crisis access points disconnected from traditional public-safety systems.

Center for American Progress & Law Enforcement Action Partnership (2020). Analysis of 911 calls in eight major cities: up to 68% could be handled without an armed officer.

National portrait of nonpolice alternative response programs (2024): 216 programs operational across the United States.

NBER Working Paper 34344: willingness-to-pay analysis finding the public places higher value on active civilian-led crisis interventions.

10. Speaking every language honestly — translation with a quality judge

InnerLight serves people in eight languages. Machine translation of safety and legal content is held to a published research standard: every translated legal-guidance card and spoken hand-off passes an independent model-judge scoring pass for semantic equivalence and natural phrasing before a person sees it, with 0.8 as the acceptance bar. Content that scores below the bar is withheld rather than shown wrong. Every new-language interface string additionally awaits native-speaker review, and that status is documented in our repository. If a device has no speech voice for the chosen language, InnerLight stays silent rather than speak the wrong language — the words remain on screen.

LLM-as-a-Judge reference-less quality estimation for machine translation, with a ≥0.8 semantic-equivalence acceptance threshold (arXiv:2503.24102).

Enomoto et al. “From LLM to NMT: Advancing Low-Resource Machine Translation with Claude” (arXiv:2404.13813) — documents where large language models still lag specialized systems on low-resource language pairs, the reason a verification pass is required.

In-context machine translation for low-resource languages (arXiv:2502.11862) — documents rare-word failure modes in underrepresented languages.

11. What the microphone reads — voice prosody

With the person’s microphone permission, InnerLight measures four classical prosodic descriptors on-device: overall vocal energy (RMS), pitch variability (autocorrelation-estimated fundamental frequency over a rolling window), speech rate (voiced-segment transitions), and a 4–8 Hz amplitude-modulation estimate. These are long-established acoustic correlates of emotional arousal in the speech-emotion literature. The raw audio for this analysis never leaves the browser; only the four numbers travel, they steer the calming music, and they are never stored. Our normalization constants are engineering estimates, labeled as such in the source code, pending calibration against a labeled recording set — consistent with our rule that nothing invented may be presented as ground truth.

Schuller B.W. (2018). “Speech emotion recognition: two decades in a nutshell, benchmarks, and ongoing trends.” Communications of the ACM, 61(5), 90–99.

Zennou H., Ouadad R., Ouhda M., Baslam M. (2026). “Real-Time Speech Emotion Recognition with a CNN-BiLSTM-Attention Deep Learning Model.” Engineering, Technology & Applied Science Research, 16(3) — the current direction of the field our simpler, on-device measures are designed to grow toward.

12. Safety signals in any language

Protective detection — recognizing when a writer may be a minor, may be substituting this tool for human connection, or may be in crisis — must not depend on the language a person writes in. In non-English sessions these signals are read by the comprehension model itself, which reads the person’s own language, rather than by English-only phrase lists. By design, an uncertain reading can only raise the level of care, never lower it, and an ambiguous statement receives a caring follow-up question rather than a verdict.

Vasan A., Stanford Brainstorm — on substitution as the primary risk of AI mental-health tools and the duty to escalate to humans (as quoted in Gold, Inside Higher Ed, 2026).

Common Sense Media — research on adolescent use of AI for mental-health advice, informing InnerLight’s heightened minor protections.

13. Holding attention through the wait — the engagement engine

Help can take fifteen minutes to an hour to arrive. InnerLight’s companion is designed to hold a person’s genuine engagement across that whole gap — rotating between comfort, small sensory activities, shared stories, and (when the person invites it) gentle humor. The activity choices follow the evidence: during acute distress the companion offers sensory and visuospatial micro-tasks, because emergency-department research shows such tasks reduce intrusive imagery in the hours after trauma, while quiz-style verbal games in that same window can make intrusions worse. Word-play and story games open up only once a person has settled. Every invitation is single, optional, and woven into conversation — never a menu.

Iyadurai L., Blackwell S.E., Meiser-Stedman R., et al. (2018). “Preventing intrusive memories after trauma via a brief intervention involving Tetris computer game play in the emergency department: a proof-of-concept randomized controlled trial.” Molecular Psychiatry, 23, 674–682.

Holmes E.A., James E.L., Kilford E.J., Deeprose C. (2010). Key finding: a visuospatial game reduced flashbacks post-trauma while a verbal quiz game did not — and in one experiment increased them. Not all engagement is beneficial engagement; activity choice must follow the evidence. PLoS ONE, 5(11): e13706.

Multilab replication (2025), Collabra: Psychology, 11(1): evidence that the visuospatial task reduces intrusions immediately — the exact window InnerLight serves — with weaker evidence for effects days later.

ANTIDOTE line of work (2024–2026): AI-guided, personally tailored imagery-competing task interventions with physiological monitoring — the research direction InnerLight’s conversational activity guidance follows.

14. The activities — matched to the state a person is in

InnerLight’s calming activities are chosen by evidence and matched to state. Research distinguishes two very different kinds of distress: high physiological arousal (racing heart, panic), where the fastest tools act directly on the body, and rumination (a looping mind), where a structured cognitive task outperforms sensory grounding. So InnerLight offers both: the physiological sigh, bilateral “butterfly” tapping, a cool-water dive-reflex reset, and an isometric wall push for the body; a categories game and counting down by sevens for the mind; plus an interactive 5-4-3-2-1 senses walk, breathing, body release, and gentle visuospatial games. Every activity is optional, guided one step at a time, and the Focus anchor never interrupts any of them.

Balban M.Y., Neri E., Kogon M.M., et al. (2023). “Brief structured respiration practices enhance mood and reduce physiological arousal.” Cell Reports Medicine, 4(1) — five minutes of cyclic sighing (double inhale, extended exhale) reduced anxiety and improved mood more than mindfulness meditation.

Ackermann S.P., et al. (2023). Meta-analysis of the diving response: moderate-to-large increases in cardiac vagal activity from cold facial stimulation — the basis of the cool-water reset and of the DBT temperature (“TIPP”) skill.

Artigas L., Jarero I. The Butterfly Hug: self-administered bilateral stimulation, developed for group trauma work and adopted within EMDR practice.

Grounding technique guidance in trauma and crisis intervention practice: proprioceptive discharge (isometric push), and the state-matching principle that cognitive tasks (categories, serial subtraction) suit rumination while somatic tools suit hyperarousal (Porges 2011, Polyvagal Theory, for the vagal mechanism).

15. Continuity of care — the wait does not end at the handoff

Reaching a provider whose next opening is weeks away moves a person’s wait rather than ending it. The evidence is unambiguous that this matters: longer waitlist lags sharply reduce the chance the first appointment ever happens, prolonged waits are associated with worsening symptoms and higher mortality, and symptoms do not usually dissipate on their own. Maintained contact during the wait is the recognized hallmark of mature programs. InnerLight therefore requires every provider to declare an honest access tier, routes to the fastest real door first, holds itself to a 72-hour network standard, and remains a person’s companion during any interim longer than that — check-ins, activities, and every crisis door — until the human is actually in the room.

Gallucci G., Swartz W., Hackerman F. (2005). Impact of the wait for an initial appointment on the rate of kept appointments at a mental health center. Psychiatric Services, 56(3), 344–346 — longer waits sharply increased first-appointment no-shows.

Systematic review, JMIR (2024): digital interventions during psychotherapy waiting lists — documents wait-associated symptom worsening, quality-of-life decline, and increased mortality; symptoms do not dissipate with time.

Carter O., et al. (2012). Increased wait-list time predicts dropout from outpatient CBT. Behaviour Research and Therapy, 50(7–8), 487–492.

Thomas K.A., et al. (2021). Systematic review of mental-health waitlist management strategies — triage, client-led approaches, and brief/single-session models as effective access interventions.

16. No dead doors — resilient handoff destinations

A crisis handoff that lands on an outage is a dead end by another name. Every critical external destination in InnerLight (legal aid, treatment finders, domestic-violence help, 211, 988 chat) is held as an ordered chain of verified alternatives; before a person is sent, the server checks that the door actually opens and routes to the first live one, logging any failover. If every door in a chain is closed the person still receives the primary target rather than nothing. Destinations that require choosing or installing an app are avoided by design: 211 opens as a website and 988 is reachable by chat as well as by phone.

Citations above reference published, peer-reviewed literature supporting the principles InnerLight applies. They do not constitute evidence that InnerLight itself is effective; that evaluation is ongoing. Full reference details are available on request.