# Ora: Behavioral Intelligence Architecture in Roveera

**Version:** 1.0 — Submission ready
**Date:** June 2026
**Target journals:** ResearchGate, SSRN, Journal of Medical Internet Research (JMIR)
**Target word count:** ~4,200 words
**DOI:** Register via Zenodo before publication

---

## Author

**Abiodun Adesina**
Founder, 3pplea Holdings LLC | Creator, Ctrl-Alt-CALM Behavioral Operating Framework and CALM Index™
Affiliation: 3pplea Holdings LLC / Roveera (roveera.com)
ORCID: [register before submission]

Abiodun Adesina is the founder of 3pplea Holdings LLC and creator of the Ctrl-Alt-CALM Behavioral Operating Framework and the CALM Index™ psychometric assessment methodology. She is the author of *Ctrl-Alt-CALM: 97 Hacks for Busy People on a Budget*, a behavioral wellness framework tested in community settings including RCCG-affiliated parish events in the United Kingdom and Nigeria. The CALM Index™ is the psychometric operationalisation of this framework. Roveera is the platform that makes it actionable. Affiliation: 3pplea Holdings LLC / Roveera (roveera.com).

---

## Abstract

Ora is the behavioral intelligence layer within the Roveera behavioral wellness platform. Unlike conversational AI interfaces designed to respond to user queries, Ora operates proactively — continuously monitoring four categories of behavioral and physiological data, identifying patterns across assessment cycles, and surfacing specific interventions without waiting for user input. This paper describes the architectural design of Ora, the four data source categories it monitors, its relationship to the CALM Index™ assessment engine, and its role in selecting and routing interventions from Roveera's behavioral intervention library. The paper argues that the key design distinction in Ora is the inversion of the standard AI interaction model: from reactive (user asks, system responds) to proactive (system observes, system recommends). The paper further describes Ora's role in the professional briefing chain — the mechanism by which Ora intelligence is transmitted to licensed human practitioners before each professional session, operationalising the connection between Stage 2 intelligence and Stage 3 human intervention. Specific architectural decisions — including rolling baseline windows, data boundary enforcement in the professional brief, adaptive anomaly routing, and the session-awareness data contract — are described with reference to the production implementation. Implications for human-AI collaborative wellness design are discussed.

**Keywords:** behavioral intelligence, proactive AI, intervention selection, behavioral wellness, CALM Index, pattern recognition, non-conversational AI, human-AI collaboration, digital mental health

---

## 1. Introduction

The dominant paradigm in AI-assisted wellness is conversational: a user presents a query or describes a current state, and the AI system responds with guidance, reflection, or resource referral. This model positions the AI as a reactive instrument — useful when engaged, invisible when not. It places the entire burden of pattern recognition on the user: the person who is least equipped to detect gradual baseline deterioration is asked to be the primary signal generator.

Ora is designed on a different premise: that the most useful behavioral intelligence does not wait to be asked.

The distinction is not merely stylistic. Behavioral research has consistently demonstrated that the populations most in need of intervention are precisely those least likely to seek it. Subclinical burnout, depression, and anxiety frequently manifest as reduced help-seeking behavior before they manifest as explicit symptom reports (Kessler et al., 2005). Users who are deteriorating tend to disengage: they check in less frequently, complete fewer habits, and avoid interactions that require reflective self-assessment. A purely reactive system — one that waits for user input — is structurally least effective when it is most needed.

Roveera's three-stage architecture addresses this by distributing intelligence across stages rather than concentrating it at the user interaction layer. Stage 1 is the CALM Index™ psychometric assessment — a validated 36-question multi-instrument battery (described separately in Adesina, 2026a) that establishes the initial behavioral baseline and seeds the live scoring system. Stage 2 is Ora — the proactive behavioral intelligence layer that continuously observes four data categories, detects deterioration patterns, and routes interventions without requiring user initiation. Stage 3 is the human professional layer — licensed therapists, coaches, and mentors who receive Ora's behavioral context brief before each session, allowing human professional skill to be directed by machine-generated intelligence.

This paper describes Stage 2: Ora's architecture, its data inputs, its pattern detection logic, and its position in the professional briefing chain.

---

## 2. The Proactive Intelligence Model

The architectural decision to build Ora as a proactive rather than reactive system has specific design consequences that distinguish it from conversational AI wellness interfaces.

A conversational AI wellness assistant operates in request-response mode: it receives user-initiated prompts and generates responses. Its knowledge of the user's behavioral state is limited to what the user explicitly reports in each interaction. It cannot observe behavioral withdrawal — the declining check-in frequency, the missed habits, the shortened sessions — because it only sees the interactions that happen, not the pattern of interactions that do not.

Ora operates differently. It receives a continuous stream of behavioral and physiological signals regardless of whether the user initiates any interaction. Behavioral withdrawal is itself a signal: the absence of expected behavior against an established baseline is as informative as presence. A user who logged mood data daily for three weeks and then went silent for four days has communicated something about their state without saying anything.

This design principle — intelligence without interface — has implications for what Ora can and cannot do. Ora does not respond to questions. It does not generate reflective prompts or conversational guidance. It does not have a chat surface. Its output is not a conversation: it is an intervention recommendation, an anomaly assessment, or a practitioner brief. Ora's presence in the platform is felt not as a conversational partner but as a continuous observational layer whose intelligence surfaces as decisions made by the system (intervention shown) or as context given to humans (practitioner brief) rather than as direct user-facing dialogue.

---

## 3. Data Architecture: Four Signal Categories

Ora monitors four categories of signal. The architecture enforces strict data boundary rules between categories — particularly at the practitioner interface — to protect privacy and prevent inappropriate data cross-referencing.

### 3.1 CALM Index™ Assessment Signals

The primary structured input. CALM Index™ composite scores, pillar-level breakdowns (Recovery, Renewal, Reach), and band classifications (Critical, Depleted, Rebuilding, Optimised) across assessment cycles provide the longitudinal baseline from which pattern detection is possible.

CALM Index™ scores are stored with timestamps in `vitals.calm_scores`. Ora's anomaly detection observes score sequences over rolling windows, identifying directional patterns rather than individual data points. A single low score is not a trigger event; a directional sequence — scores declining across two or more consecutive cycles — is. This approach reduces false positive alert rates and focuses Ora's intervention activity on persistent behavioral patterns rather than transient variance.

Band transitions are primary trigger conditions. Downward band movement (Rebuilding → Depleted, Depleted → Critical) activates the anomaly assessment pipeline. The magnitude threshold for trend detection is a ±3-point delta across the observation window: movements smaller than this are classified as stable.

For the practitioner briefing system, CALM Index™ data is handled under a strict abstraction policy: raw numeric scores are never transmitted to the professional briefing model. The practitioner receives only directional labels ("improving," "declining," "stable") and band classifications ("Rebuilding," "Depleted") — preserving clinical utility while preventing over-reliance on specific numerical values that carry more precision than the underlying measurement warrants for practitioner-level decision-making.

### 3.2 Wearable Physiological Signals

Where users have connected wearable devices, Ora receives nine physiological signal streams: HRV (RMSSD), resting heart rate (bpm), respiratory rate (bpm), sleep duration (minutes), sleep efficiency (percentage), SpO₂ (percentage), active calories, energy expenditure (kcal), and skin temperature deviation (Celsius).

Each signal has a dedicated rolling baseline window calibrated to the signal's characteristic variability: HRV, resting HR, respiratory rate, and SpO₂ are computed over 21-day windows; sleep duration and efficiency over 14-day windows; activity signals over 7-day windows. These windows reflect the time constants of the underlying physiological processes: HRV baseline takes several weeks to stabilise; sleep duration is more variable week-to-week; activity patterns reflect shorter behavioral cycles.

Where a user's personal standard deviation for a signal is zero (insufficient variance in the observation window), population-level standard deviation fallbacks are applied: HRV RMSSD ± 15ms, resting HR ± 8bpm, sleep duration ± 60min, SpO₂ ± 1.5 percentage points. This ensures that anomaly detection remains functional during the early user lifecycle before a robust personal baseline has been established.

Physiological data is treated as a leading indicator for CALM Index™ pillar movement. HRV decline and sleep disruption reliably precede self-reported mood and stress deterioration by 2–3 days in the research literature (Plews et al., 2013; Åkerstedt & Gillberg, 1981). Ora's architecture is designed to allow physiological signals to trigger early intervention ahead of assessment score movement, reducing the detection latency inherent in periodic self-report assessment cycles.

The data gap condition is defined as: wearable unsynced for more than 48 hours AND fewer than two mood log entries in the last 7 days. When this condition is true, Ora's anomaly assessments are flagged as low-confidence and the intervention routing accounts for data uncertainty.

### 3.3 Behavioral Engagement Signals

Check-in frequency, session engagement, habit completion rates, app interaction patterns, and breathing session records constitute the behavioral engagement layer. These signals are observational: they do not require the user to report any explicit state. They are generated as side effects of normal platform use.

The behavioral engagement layer captures the withdrawal pattern that purely assessment-based systems miss. The behavioral signals that most reliably precede score deterioration are: declining check-in frequency, reduced habit completion rate, shortened or abandoned breathing sessions, and absence from community engagement. These patterns do not require the user to report feeling worse — they emerge from the absence or attenuation of previously established behavioral routines.

Breathing session data contributes a particularly valuable physiological signal: `vitals.breathing_sessions.hrv_after` records HRV measured after breathing exercise completion, providing a non-wearable source of physiological data for users without connected devices.

### 3.4 Session and Intervention Signals

The feedback loop layer. Professional session history (session type, timing, CALM delta direction post-session, pending task state, task completion rate) informs Ora's understanding of which interventions have produced measurable score change and whether professional session engagement is being followed through.

Session type carries specific pillar implications: therapy sessions are mapped to Recovery, coaching sessions to Renewal, and mentoring sessions to Reach. This mapping preserves the alignment between the type of professional support being engaged and the pillar that type of support is theoretically designed to address.

Post-session CALM delta direction — the direction of CALM Index™ score movement after a session compared to the score immediately before — provides a session-level effectiveness signal. Ora surfaces this signal (as directional label only, never as raw number) in the practitioner prep brief to allow the professional to calibrate session approach based on what has produced improvement versus what has not.

Task completion rate from previous sessions (completed tasks / total assigned tasks) provides a measure of between-session follow-through. Low completion rates are associated with reduced long-term engagement and are surfaced in practitioner briefs to inform session planning.

The session signals are governed by a strict data boundary protocol (see Section 5):
- ✅ Session type (therapy | coaching | mentoring)
- ✅ Pillar impacted (derived from session type mapping)
- ✅ Session timing
- ✅ CALM delta direction (directional only, no raw scores)
- ✅ Whether pending tasks exist
- ✅ Task completion rate (fraction)
- ❌ Q1/Q2/Q3/Q4 session evaluation values (excluded from Ora session context)
- ❌ Practitioner identity (name, ID, type)
- ❌ Session notes or concern text
- ❌ Free-text feedback

The window for session signals is 30 days, consistent with the behavioral engagement signal window.

---

## 4. Pattern Recognition and Anomaly Detection

Ora identifies three classes of behavioral pattern that trigger intervention routing:

**Band deterioration pattern.** Composite CALM Index™ score moving toward lower bands across consecutive assessment cycles. Primary trigger condition for the anomaly assessment pipeline. The anomaly assessment (described below) classifies the deterioration and routes to one of four actions.

**Pillar-specific drift.** Targeted pillar decline (e.g., Recovery deteriorating while Renewal and Reach remain stable) suggesting a domain-specific intervention need. Pillar-specific drift allows Ora to surface targeted interventions from the appropriate pillar toolkit rather than a general wellness response.

**Behavioral withdrawal pattern.** Reduced check-in frequency, session avoidance, declining habit completion, or absence from engagement that has been consistent for 2+ weeks. This pattern class is particularly important because it often precedes detectable score deterioration and represents an opportunity for early intervention before the assessment-layer signal appears.

When a band deterioration pattern or significant anomaly is detected, the anomaly assessment pipeline activates. The assessment runs under strict safety protocols: it always uses the highest available model tier (never downgraded for cost) and writes every invocation immutably to the audit log regardless of outcome.

The anomaly assessment produces a recommended action from four categories:

| Action | Description |
|--------|-------------|
| `coach_nudge` | Write intervention context to Redis (1h TTL); Ora surfaces a behavioral intervention or check-in prompt |
| `crisis_resources` | Trigger safety module; suppress gamification elements; surface crisis support resources |
| `practitioner_alert` | Write immutable alert to platform audit log and Redis flag; escalate to professional layer |
| `monitor` | Log only; no user-facing action; continue observation |

Crisis routing (`crisis_resources`) is triggered when anomaly signals suggest acute distress or safety concern. The gamification suppression is an important architectural detail: wellness streaks and achievement markers are contextually inappropriate for users in crisis, and their suppression is an active safety design decision.

---

## 5. Intervention Selection

When a pattern is identified and a trigger condition met, Ora selects a specific behavioral intervention from Roveera's behavioral intervention library — the curated collection of hacks, habits, affirmations, challenges, and exercises sourced from the Ctrl-Alt-CALM Behavioral Operating Framework.

Selection is determined jointly by the user's current CALM Index™ pillar scores, their established primary pillar assignment, the detected pattern class (band deterioration, pillar-specific drift, or behavioral withdrawal), and their historical intervention engagement record.

The behavioral intervention library is organized into 13 toolkit categories aligned to the three CALM Index™ pillars:

**Recovery toolkits (5):** Crisis Calming, Anxiety Toolkit, Stress Release, Emotional Processing, Burnout Recovery.

**Renewal toolkits (4):** Sleep Reset, Energy Restoration, Movement & Vitality, Nutrition & Nourishment.

**Reach toolkits (4):** Goals & Performance, Purpose & Meaning, Relationships & Connection, Mindset Reset.

Primary toolkit prescription follows the user's assigned primary pillar. Emergency overrides exist for safety conditions: if the user's primary pillar is Recovery and their band is Critical, the toolkit prescription is overridden to Crisis Calming (toolkit 1) regardless of the otherwise-highest-need toolkit. Similarly, ISI-7 scores indicating moderate or severe clinical insomnia (raw ISI-7 total ≥ 15) trigger Sleep Reset inclusion regardless of pillar assignment, reflecting the primacy of sleep disruption as a cross-pillar recovery barrier.

For new users (cold-start condition), toolkit selection defaults to the pillar-specific prescription catalogue without engagement-history weighting. As a user's engagement history accumulates, the selection progressively weights interventions that have historically produced higher engagement rates for the user's current pillar and band profile.

Frequency limits are applied to prevent intervention fatigue: the same toolkit category is not surfaced on consecutive days, and crisis-level interventions are not suppressed by frequency limits regardless of recency.

---

## 6. The Professional Briefing System

The practitioner preparation brief is the mechanism by which Ora's intelligence reaches the human professional layer. Before each scheduled session with a licensed therapist, coach, or mentor, Ora generates a concise pre-session brief accessible to the practitioner through the platform dashboard.

The brief contains four structured fields:
1. **Context summary** (2–3 sentences): primary pillar context, wellness trend direction, relevant session history. Written in warm, collegial language — not clinical jargon.
2. **Task follow-up** (1 sentence or null): status of any pending tasks from the practitioner's previous sessions with this client.
3. **Suggested focus areas** (exactly 2): specific and actionable areas for session attention, derived from the behavioral and pillar context.
4. **Heads-up** (1 sentence or null): a clinical flag if genuinely warranted — typically triggered by declining trend combined with low task follow-through.

The brief is cached in Redis with a 24-hour TTL keyed to the session ID. First sessions return a template brief (no AI inference) since insufficient history exists to generate a meaningful behavioral picture.

**Data boundary enforcement** in the professional brief is a critical architectural decision. The brief is generated via a language model call; the language model receives only strictly filtered inputs:

| Input permitted | Input excluded |
|----------------|----------------|
| Primary pillar (label only) | Raw CALM Index™ scores |
| Wellness band (label only) | PHQ-9, GAD-7, PSS, ISI, SWLS, BAT scores |
| Wellness trend direction | Coach/Ora interaction transcripts |
| Session intention (client-stated, from booking) | Other practitioners' sessions or observations |
| Previous sessions with THIS practitioner only | Community posts or personal communications |
| Q1 session rating from last session with this practitioner | Session notes from any practitioner |
| Pending and completed tasks (this practitioner only) | |
| Domain focus summaries from this practitioner's previous sessions | |

This boundary serves two purposes. First, it protects client privacy by ensuring that information from one professional relationship is not disclosed to a different professional without explicit consent. Second, it prevents the practitioner from being anchored to specific numerical values — the directional labels ("declining," "Depleted band") convey clinically useful information without introducing spurious numerical precision into the practitioner's assessment.

The brief supports multilingual output: the practitioner receives the brief in their registered language preference (English, French, or Nigerian Pidgin), with the tone calibrated for collegial rather than clinical communication.

---

## 7. Discussion

### 7.1 Proactive vs. Reactive AI in Behavioral Wellness

The proactive architecture described in this paper is one instance of a broader design question for AI-assisted health systems: where should the intelligence boundary sit? In reactive systems, the intelligence boundary is at the user's expressed need — the system acts when asked. In proactive systems, the intelligence boundary is at the system's observation capability — the system acts when it detects a need.

The proactive model carries specific advantages in the behavioral wellness domain. First, subclinical deterioration frequently presents as behavioral withdrawal before it presents as explicit symptom report — a purely reactive system will miss the early signal. Second, the populations most in need of intervention are systematically less likely to initiate help-seeking (Clement et al., 2015); proactive detection partially offsets this structural access barrier. Third, the integration of physiological signals with behavioral and assessment data allows leading-indicator detection ahead of lagging self-report data, reducing the latency between emerging deterioration and appropriate response.

The proactive model also carries specific risks. False positives — Ora triggering intervention when no deterioration is actually occurring — carry a real cost: inappropriate escalation, user annoyance, and erosion of trust in the system's signal reliability. The four-level action routing (coach nudge → practitioner alert → crisis resources → monitor) is designed to calibrate the response magnitude to the confidence and severity of the detected pattern. The lowest-confidence or lowest-severity detections route to `monitor` (no user-facing action); only high-confidence, high-severity patterns route to crisis resources or practitioner alerts.

### 7.2 Intelligence Without Interface

A frequently encountered assumption in AI wellness design is that a more capable AI system should have more visible interface — more conversational, more responsive, more present. Ora's architecture inverts this assumption: the intelligence layer has no dedicated interface. Its outputs are experienced by the user as interventions surfaced by the platform, not as Ora recommendations, and by practitioners as a brief that arrives before their session, not as an AI consultation.

This design is deliberate. The goal is not for Ora to be present as an agent in the user's experience; it is for the user's experience to be better because Ora is present. The distinction has practical implications for transparency: users are informed that behavioral intelligence monitors their patterns and informs interventions and practitioner briefs, but the interaction model is not built around that awareness.

### 7.3 Limitations

**Cold-start problem.** Ora requires a minimum behavioral history before pattern detection is meaningful. New users have no baseline against which to detect drift; population-level physiological fallbacks (Section 3.2) partially address this but cannot substitute for individual baseline establishment. The assessment seed system (described in Adesina, 2026a) ensures that a scoring baseline exists from Day 1, but pattern detection sensitivity improves progressively with 30+ days of behavioral data.

**Wearable data gaps.** Not all users have connected wearable devices, and among those who do, data continuity is variable. The data gap condition (wearable unsynced >48h AND mood absent <2 entries/7 days) is a conservative definition; real-world data quality varies significantly. Ora's architecture must degrade gracefully when physiological leading-indicator data is absent, falling back to assessment-cycle and behavioral engagement signals alone.

**Engagement signal sparsity.** Intervention dismissal — a user being shown an intervention and not engaging — may reflect multiple states: irrelevance, poor timing, intervention format mismatch, or genuine disinterest. Distinguishing these states from dismissal data alone is not reliably possible. Ora's engagement-history weighting system accumulates dismissal patterns and progressively de-weights dismissed intervention types, but the signal is inherently noisy.

**Explainability.** Proactive recommendations need justification pathways for users: when Ora surfaces an intervention, users benefit from understanding why. The current architecture provides basic pillar-level rationale ("This aligns with your Recovery pillar focus area") but does not surface the full pattern evidence that triggered the recommendation. Improved intervention rationale presentation is identified as a priority for future development.

**Professional brief limitations.** The practitioner prep brief is useful for established relationships but uninformative for first sessions (returning a template placeholder). The strict data boundary — while privacy-protective — also limits the clinical depth of the brief. A practitioner may benefit from knowing that a client's ISI-7 scores indicate clinical insomnia without knowing the specific PHQ-9 total; the current architecture either shares all assessment data or none. A more granular data boundary architecture that allows specific clinical signals to be shared with appropriate consent is an identified enhancement.

### 7.4 Future Research

**Prospective effectiveness study.** A controlled comparison of CALM Index™ band movement at 90 days in users with active Ora engagement versus users in an assessment-only condition would provide the first prospective evidence for Ora's behavioral intelligence contribution to wellness outcomes.

**Intervention type specificity.** Which toolkit categories produce the greatest band movement for each pillar and band profile? The intervention library is currently selected based on theoretical pillar alignment; empirical evidence for differential effectiveness by intervention type, user profile, and presenting band would allow the selection algorithm to be refined from theory-based to evidence-based.

**Professional briefing impact.** A comparative study of professional session outcomes with versus without access to the Ora prep brief would quantify the contribution of the intelligence layer to the human professional session — specifically whether practitioner access to directional CALM and pillar data improves the convergence between session focus and the client's primary behavioral deficit domain.

**Proactive vs. reactive mode comparison.** A within-platform comparison of user outcomes under proactive Ora intervention (standard) versus reactive-only mode (interventions surfaced only on user request) would provide direct evidence for the behavioral advantage of the proactive architecture.

---

## References

Åkerstedt, T., & Gillberg, M. (1981). The circadian variation of experimentally displaced sleep. *Sleep*, 4(2), 159–169.

Adesina, A. (2026a). The CALM Index™: A three-pillar psychometric framework for behavioral wellness. 3pplea Holdings LLC / Roveera. https://roveera.com/research

Adesina, A. (2023). *Ctrl-Alt-CALM: 97 Hacks for Busy People on a Budget*. 3pplea Holdings LLC.

Clement, S., Schauman, O., Graham, T., Maggioni, F., Evans-Lacko, S., Bezborodovs, N., … Thornicroft, G. (2015). What is the impact of mental health-related stigma on help-seeking? A systematic review of quantitative and qualitative studies. *Psychological Medicine*, 45(1), 11–27.

Ghandeharioun, A., Fedor, S., Sangermano, L., Ionescu, D., Alpert, J., Dale, C., … Picard, R. (2017). Objective assessment of depressive symptoms with machine learning and wearable sensors. *Proceedings of the 7th International Conference on Affective Computing and Intelligent Interaction (ACII)*, 325–332.

Insel, T. R. (2017). Digital phenotyping: Technology for a new science of behavior. *JAMA*, 318(13), 1215–1216.

Jiang, F., Jiang, Y., Zhi, H., Dong, Y., Li, H., Ma, S., … Wang, Y. (2017). Artificial intelligence in healthcare: Past, present, and future. *Stroke and Vascular Neurology*, 2(4), 230–243.

Kessler, R. C., Berglund, P., Demler, O., Jin, R., Merikangas, K. R., & Walters, E. E. (2005). Lifetime prevalence and age-of-onset distributions of DSM-IV disorders in the National Comorbidity Survey Replication. *Archives of General Psychiatry*, 62(6), 593–602.

Mohr, D. C., Zhang, M., & Schueller, S. M. (2017). Personal sensing: Understanding mental health using ubiquitous sensors and machine learning. *Annual Review of Clinical Psychology*, 13, 23–47.

Plews, D. J., Laursen, P. B., Stanley, J., Kilding, A. E., & Buchheit, M. (2013). Training adaptation and heart rate variability in elite endurance athletes: Opening the door to effective monitoring. *Sports Medicine*, 43(9), 773–791.

Satyanarayanan, M., Bahl, P., Cáceres, R., & Davies, N. (2001). The case for VM-based cloudlets in mobile computing. *IEEE Pervasive Computing*, 8(4), 14–23.

Torous, J., Kiang, M. V., Lorme, J., & Onnela, J. P. (2016). New tools for new research in psychiatry: A scalable and customizable platform to empower data driven smartphone research. *JMIR Mental Health*, 3(2), e16.

Topol, E. J. (2019). High-performance medicine: The convergence of human and artificial intelligence. *Nature Medicine*, 25(1), 44–56.

World Health Organization. (2004). *Promoting mental health: Concepts, emerging evidence, practice*. WHO Press.

---

*© 2026 3pplea Holdings LLC. This work is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). To view a copy of this licence, visit https://creativecommons.org/licenses/by/4.0/*

*Cite as: Adesina, A. (2026). Ora: Behavioral intelligence architecture in Roveera. 3pplea Holdings LLC / Roveera. https://roveera.com/research*