Tecnologia / Analisis profundo crossmodal y deteccion de congruencia

ORACLE

Raw data tells you what happened. Patterns tell you what it means.

CYGNUS extracts signals from three channels: the face (Action Units, spatial landmarks), the voice (pitch, rate, volume, pauses, vocal quality), and the body (head position, torso orientation, gestural activity). In current video-call deployments, that body channel usually means the upper torso through roughly the chest or half-upper-body frame, because that is what most camera setups actually provide. The architecture can extend further as input conditions expand. Even in its current form, CYGNUS produces hundreds of numerical values per second, each describing one specific observable movement or measurement. This data stream is rich, precise, and completely uninterpreted.

ORACLE is the model that makes these numbers speak. It is the deep analysis and intra-session pattern recognition engine of the OPM pipeline. No other model analyzes as deeply. Where CYGNUS sees individual measurements, ORACLE sees relationships. Where CYGNUS records values, ORACLE detects convergence, divergence, and anomaly.

Think of ORACLE as the part of the brain that notices: something just shifted. A trained therapist who suddenly senses that the patient's words and body language are not telling the same story. A negotiator who picks up on a micro-hesitation that nobody else caught. That intuitive recognition of pattern is exactly what ORACLE does, except it does it numerically, consistently, and across all available channels at once.

Listen Mode

A guided audio version of this page, adapted for clarity and flow rather than read word for word.

0:00/ 0:00
Loading
Curated narration
Model Character

ORACLE is the pattern-recognition model of the OPM system. It does not invent meaning out of nothing. It evaluates relationships, recurrence, timing, and cross-channel agreement inside the signal stream.

Rule-based engine
Validated behavioral patterns encoded as evaluable rules
Cross-modal field
Face, voice, and body referenced against one another
Intra-session memory
Recurrence, rhythm breaks, and turning points inside one session
Structured findings
Confidence, category, timestamp, duration, recurrence
How Pattern Recognition Works

Raw measurements become findings when relationships emerge.

ORACLE operates on a rule-based engine. Each rule encodes a specific behavioral pattern observed in research and validated against real interaction data. These are not simple threshold checks. They are multi-signal conditions that consider combinations, timing, duration, and cross-channel relationships.

A single ORACLE rule might look like this in plain language: "If vocal pitch is rising while the lip corners are pulling downward and the shoulders are elevating, and this combination persists for more than two seconds, flag a cross-modal divergence between vocal tone and facial or postural signals."

That rule checks three channels simultaneously, considers timing, and produces a structured finding. ORACLE runs its full rule library in parallel, continuously evaluating the incoming CYGNUS stream.

The Rule Structure

Every ORACLE rule follows the same architecture.

Trigger Conditions

What combination of signals must be present? These conditions can span all three CYGNUS channels and require specific thresholds, directional changes, and multi-signal co-occurrence.

Temporal Requirements

How long must the pattern persist? Some patterns matter because they are brief. Others matter only because they last. ORACLE encodes that distinction directly into the rule.

Confidence Scoring

Each finding carries a confidence score from 0 to 1. ORACLE does not emit simple yes or no outputs. It reflects how closely the observed signals matched the rule.

Finding Classification

Each match is categorized structurally: convergence, divergence, escalation, de-escalation, anomaly, or shift.

Example Rule Flow
1
Trigger
Pitch rising + lip corners downward + shoulders elevating
2
Timing
Pattern persists beyond 2 seconds
3
Scoring
Observed stream matches rule at confidence 0.87
4
Finding
Cross-modal divergence emitted as structured finding
What ORACLE Looks For

The major categories of behavioral patterning.

Cross-Modal Convergence

When multiple channels tell the same story. If the voice is steady, the face is relaxed, and the body is open, these signals converge. Convergence is itself meaningful because it indicates that the observable signals are internally consistent.

Cross-Modal Divergence

When channels contradict each other. A voice that sounds calm while the face shows tension and the body contracts. This is one of ORACLE's most valuable capabilities because divergence is extremely difficult for humans to spot in real time.

Signal Escalation

When a value or combination of values is increasing over time: rising pitch, increasing brow-lowering activation, slowly contracting posture. Individually these may be subtle. Together they become a clear escalation pattern.

Signal De-escalation

The inverse. Values that were elevated are settling. Tension that was building is releasing. ORACLE tracks this trajectory with the same precision as escalation.

Hesitation Clusters

A pause in speech, a lip press, a gaze aversion, a postural freeze. When these co-occur inside a tight temporal window, ORACLE flags the cluster.

Rhythm Breaks / Temporal Asymmetry

Disruptions to established behavioral rhythm or unexpected timing misalignment across channels. These are often nearly impossible to hold in view manually during live interaction.

Comparison Frame

Single-channel attention vs cross-modal analysis

Human expert without ORACLE

Strong intuition, limited parallel tracking, attention usually anchored to one or two channels at a time, difficult to hold recurrence and timing asymmetry in view across an entire session.

ORACLE

Evaluates face, voice, and body in parallel, tracks recurrence inside the current session, scores confidence continuously, and checks temporal relationships automatically against the full rule library.

Human expertORACLE
1.0
0.8
0.6
0.4
0.2
usable pattern depth
cuedivergencetimingrecurrencesession
The Cross-Modal Advantage

The combination tells a different story.

Most perception systems analyze one channel at a time. A facial system reads the face. A voice system reads the voice. They operate in isolation.

ORACLE's fundamental design principle is cross-modal analysis: every finding considers data from multiple channels simultaneously. This matters because human communication is inherently cross-modal. A smile means something different when it is accompanied by a steady voice and open posture than when it is accompanied by a shaking voice and contracted shoulders.

ORACLE holds all three channels in its field of view at all times. When it detects a facial pattern, it immediately checks what the voice and body were doing at that same moment. When it detects a vocal shift, it checks the face and posture. This cross-referencing happens automatically, on every rule evaluation, for every second of the interaction.

The result is a perception depth that cannot be achieved by stacking single-channel systems. It is not about having more data. It is about understanding how the data relates across channels.

Intra-Session Pattern Detection

ORACLE remembers the current session while it analyzes it.

ORACLE does not just evaluate the current moment. It maintains awareness of everything that has happened within the current session.

If a specific divergence pattern appeared at minute 3 and appears again at minute 17, ORACLE recognizes the recurrence. If hesitation clusters consistently appear when a particular topic is discussed, ORACLE flags the correlation. If an escalation pattern that was building for 10 minutes suddenly reverses, ORACLE notes the turning point.

These intra-session patterns are distinct from the cross-session tracking that TRACE handles. ORACLE's scope is the current session, from beginning to end. Within that scope, it builds a running model of the behavioral landscape and continuously evaluates new data against what it has already seen.

A finding that says "divergence detected at t=14:32" is useful. A finding that says "divergence detected at t=14:32, consistent with similar divergence at t=3:18 and t=8:45, all co-occurring with topic X" is dramatically more useful.

Session Timeline
0.90
0.75
0.60
0.45
3:18
8:45
14:32
Topic X
confidence / correlation
Divergence0.58Hesitation0.64Recurring divergence0.77Correlation0.88
recurring pattern strength across session
Structured Findings

What ORACLE produces

ORACLE's output is a structured set of findings. Each finding describes the rule that triggered, when it appeared, which channels contributed, how strongly the data matched, how long the pattern persisted, and whether it recurred earlier in the same session.

Field
Description
Example
Rule ID
Which rule triggered
CROSS_MODAL_DIVERGENCE_07
Timestamp
When in the session
t = 842.3s
Channels Involved
Which CYGNUS channels contributed
Facial + Vocal
Confidence
How strongly the data matched the rule
0.87
Category
Structural classification
Divergence
Signal Summary
Specific values that triggered the rule
AU4: 0.72, AU15: 0.41, Pitch: +23Hz from baseline
Duration
How long the pattern persisted
3.2 seconds
Recurrence
Whether the pattern appeared earlier
Similar pattern at t=218.1s, t=507.6s
What ORACLE Doesn't Do

Observable patterns are not verdicts.

ORACLE's rules are explicitly behavioral. They detect observable patterns in measurable signals. There are important things ORACLE avoids by design.

ORACLE does not label emotions. It will not tell you someone is happy, angry, or nervous. It tells you that specific Action Units are activated at specific intensities while specific vocal features are present and specific postural parameters are shifting.

ORACLE does not assess truthfulness. Cross-modal divergence is an observable phenomenon, not a lie detector. When channels disagree, ORACLE reports the disagreement.

ORACLE does not make predictions. It analyzes what has happened and what is happening. It does not project what will happen next.

ORACLE does not compare people to each other. All analysis is individual. Population-level baselines exist in the rule library, but the findings are always about the person in front of the system.

The Rule Library

Two rulesets, different inputs.

ORACLE's rules are organized into two major rulesets, each designed for a different input configuration.

Video Rules (Cross-Modal Rules)

When CYGNUS Standard or CYGNUS Lite is active, ORACLE runs the video ruleset. These rules evaluate facial, vocal, and postural signals simultaneously.

They cover facial convergence patterns, postural shifts that correlate with facial changes, temporal relationships between facial and body movement, gestural patterns that accompany or contradict facial expressions, and baseline deviations in facial and postural parameters.

The video ruleset currently contains 47 validated cross-modal rules. That number continues to grow as new patterns are validated through research and deployment.

ECHO Rules (Audio)

When CYGNUS ECHO is active, ORACLE switches to the ECHO ruleset. These rules are purpose-built for vocal-only analysis and evaluate signals exclusively from the audio channel.

They cover prosodic patterns, pause architecture, vocal quality changes, temporal relationships between vocal features, emotional escalation curves, topic-bound hesitation patterns, and speech fluency dynamics.

The ECHO ruleset currently contains 34 validated rules. These are not simplified video rules. They are a dedicated ruleset designed specifically for what voice alone can reveal.

ORACLE in the OPM Pipeline

Second position. Deep analysis. Findings forward.

ORACLE sits at the second position in the pipeline, receiving data directly from CYGNUS and passing findings forward to LUCID.

Its input is CYGNUS's output: the continuous multi-channel numerical stream. Its output is a structured set of findings describing detected patterns, confidence levels, temporal characteristics, and recurrence inside the session.

ORACLE does not modify CYGNUS data. It consumes it, evaluates it against its rule library, and produces findings. The original signal data continues downstream alongside the findings so that LUCID and TRACE can access both the raw measurements and ORACLE's analysis.

Frequently Asked Questions

Common ORACLE questions

How many rules does ORACLE use?+

The video ruleset currently contains 47 cross-modal rules. The ECHO ruleset currently contains 34 audio-only rules. Both are in active development and continue to grow as new patterns are validated.

Can ORACLE detect deception?+

No. ORACLE detects cross-modal divergence, not deception. Divergence can arise for many reasons: discomfort, uncertainty, cognitive load, social performance, or cultural expression norms.

Does ORACLE learn or adapt during a session?+

The rules themselves are fixed for a given deployment. What changes over the course of the session is context: ORACLE accumulates awareness of recurring patterns and behavioral shifts.

Is ORACLE real-time?+

Not as ORACLE alone. Real-time deployments are handled by ORACLE RT, which was built specifically for that mode in combination with CYGNUS Lite. Standard ORACLE operates in near-real-time or batch-oriented configurations depending on deployment context.

Can ORACLE operate without all three channels?+

Yes, but the mode matters. When CYGNUS ECHO is active, ORACLE can run the audio ruleset without the video and full cross-modal rules. In real-time audio-first or reduced-channel contexts, ORACLE RT is the system built specifically for that operating mode. The findings remain useful, but narrower in scope than full three-channel analysis.

What makes ORACLE different from traditional behavioral analysis software?+

Traditional tools often analyze one channel in isolation and produce labels. ORACLE analyzes all available channels simultaneously, tracks intra-session patterns, and produces structured findings with confidence scores instead of categorical emotional outputs.

Does ORACLE replace human expertise?+

No. ORACLE increases perceptual consistency and depth, but it does not replace clinical, therapeutic, investigative, or coaching judgment. It exists to extend human perception and move closer to the broader goal of extended humans, not to remove the human expert from the loop. It produces findings, not verdicts.

What happens if only two channels are available?+

ORACLE can still produce valuable findings when at least two channels are active. The deepest cross-modal rules depend on multiple channels, but the system degrades by scope, not by turning into labels.

Can institutions audit which ruleset is active?+

Yes. Institutional deployments can be technically reviewed to verify whether ORACLE is running the video ruleset, the ECHO ruleset, or a narrower channel configuration.

Technical Summary

ORACLE at a glance

Input
CYGNUS multi-channel signal stream
Output
Structured pattern findings with confidence scores
Analysis Type
Rule-based, cross-modal, temporal
Rule Categories
Video Ruleset (47), ECHO Ruleset (34)
Finding Types
Convergence, Divergence, Escalation, De-escalation, Hesitation Cluster, Rhythm Break, Temporal Asymmetry
Confidence Range
0.0 to 1.0 (continuous)
Scope
Intra-session (current session only)
Real-Time Capable
Yes, with CYGNUS Lite
Emotion Labels
None produced
Downstream

ORACLE turns measurement into structured findings. The next model, LUCID, places those findings into human-readable contextual interpretation.