Preprint methodology · empirical validation in preparation

The science
behind the
mirror.

FaceMirror is not built on intuition or prompt engineering. It is built on a preprint-stage methodological framework, grounded in peer-reviewed literature, and a triangulation architecture that has no shortcut.

Preprint submitted · PsyArXiv 2026
2 patents pending · 3 independent channels
210+ dynamic archetype descriptors
Scroll
01 · The Problem

Why most behavioral reports
feel accurate but reveal nothing.

Automated personality assessments routinely score high on user-rated accuracy. Users read them and confirm they fit. Yet most of these reports are informationally inert — they tell people only what they already know about themselves. We call this the plausible-but-flat problem.

01.1

The Barnum Effect

Named after P. T. Barnum, this well-documented phenomenon describes how generic statements feel personally accurate to nearly anyone who reads them. Most AI-generated profiles achieve their accuracy by remaining at a level of generality that says nothing of substance.

01.2

Personalization ≠ Insight

A report can include your name, your profession, and your actual phrases — and still show you nothing new. Technical personalization does not produce the moment of recognition a genuine self-reflection tool must deliver.

01.3

Single-channel blindness

Any single observation method — whether self-report, voice, or facial analysis alone — captures only one perspective. The disagreements between channels are invisible from inside any one of them. That is exactly where the insight lives.

01.4

What insight actually requires

Genuine self-reflection requires the user to encounter a pattern they can verify in their own experience but had not previously articulated. This can only come from observable structure the user has no ordinary way to see — not from interpretation alone.

"A description that says you sometimes prefer to act on intuition, but at other times you want to plan carefully will be confirmed by nearly anyone. It is true; it is also useless."
Stolyarov et al. (2026) · Recognition + Shift · §1.1
02 · The Framework

Recognition + Shift:
a two-layer architecture.

The core design principle specifies that every report must contain two structurally distinct layers — produced from different sources, answering different questions, and carrying different evidentiary status. Collapsing them is a methodological error.

Layer I

Recognition

Patterns the user can verify in themselves through ordinary self-observation. This layer establishes that the system understands the user. Without it, the user has no basis to trust the second layer.


  • Derived from self-report and confirmed behavioral signals
  • Credibility floor — not insight in itself
  • Designed for test-retest reliability
  • Mirrors the trait dimension of personality architecture
Layer II

Shift

Patterns visible to the system through cross-channel divergence — not available to the user from any single channel of self-observation. This is the only place where the report claims to show the user something they cannot already see.


  • Strictly observation-grounded — traceable to specific inter-channel divergence
  • Not interpretation — observable structural regularities
  • Expected to be more state-sensitive across sessions
  • Mirrors the state dimension of personality architecture
INPUT LAYER Self-Report 44-item instrument Verbal Association rhythmic stimulus Facial Expression continuous capture CROSS-CHANNEL DIVERGENCE detection + localization OUTPUT LAYER RECOGNITION confirmed patterns high retest reliability SHIFT divergence-derived state-sensitive Executive Profile + ARCHETYPE DESCRIPTOR
Figure 1 Recognition + Shift framework architecture. Three independent input channels feed a cross-channel divergence detection stage. Output is split into two structurally distinct layers with different evidentiary status.

Methodological commitment: The two-layer separation is not stylistic. A reader who treats Recognition as evidence for Shift — or Shift as a more vivid form of Recognition — misuses the framework. The separation is the honesty claim the entire system rests on. Divergences are not interpretations; they are observable cross-channel patterns — detectable by a multimodal system, invisible to the user from any single channel.

03 · Methodology

Divergence is the method.
Three channels make it visible.

Most behavioral tools measure something and report it. FaceMirror's logic is different: we are not primarily interested in what each channel says. We are interested in where the channels disagree — and why that disagreement is invisible to the person being observed from any single channel alone. Three independent inputs are the minimum required to detect divergence and reason about which channel is anomalous. Two is not enough.

01

Free Verbal
Association

One word per beat of a metronome for up to five minutes. The rhythmic constraint reduces volitional control over content selection. Statistical properties of the output — not semantic content alone — are extracted and compared against the self-report.

Input channel 1 · Behavioral output
02

Facial Expression
Analysis

Facial-expression features extracted via webcam during the same session. Provides a second comparison axis independent of both self-report and speech. When it agrees with one of the first two channels, the disagreeing channel becomes interpretable.

Input channel 2 · Nonverbal signal
03

Self-Report
Instrument

A standardized 44-item psychometric instrument with established validity evidence in the literature. Captures how the user perceives and presents themselves — the self-model. Subject to social desirability and limited self-knowledge by design. Its value is not in what it measures alone, but in what it disagrees with.

Input channel 3 · Self-model

The channels exist to produce disagreements. The disagreements are the finding.

When any two channels diverge, FaceMirror does not average them or choose the "more reliable" one. The divergence itself — observable, traceable, specific — is surfaced to the user as content. This is the architectural commitment that separates FaceMirror from any single-channel or aggregation-based system: we built the instrument specifically to detect and use divergence, not to resolve it away.

divergence axis divergence axis divergence axis Verbal Association channel 2 Self-Report Instrument channel 1 Facial Expression channel 3 SHIFT ZONE 2 channels: detect cannot localize 3 channels: detect + localize anomaly
Figure 2 Three-channel triangulation system. Each edge is a comparison axis. The Shift zone is accessible only when all three channels are present. Two channels detect disagreement; three channels localize which channel is anomalous.
04 · Archetype Model

State-trait architecture.
210+ computed descriptors.

Each report includes a two-word archetype label — a structural noun paired with a modal participle. This decomposition mirrors the state-trait distinction in classical personality research (Spielberger et al., 1970) and yields a predicted, empirically testable pattern of test-retest reliability.

The structural noun is derived deterministically from quantitative profile dimensions — a stable functional role that should replicate across sessions. The modal participle is generated from session-specific signals and is expected to vary — it describes a behavioral mode, not a permanent characteristic.

Unlike fixed typologies with 8 or 16 categories, the descriptor space is computed from a vocabulary matrix of 15 participles × 14 nouns. No two users receive identical archetypes unless their profiles are quantitatively equivalent.

Structural noun

Trait layer

Stable functional role. Derived algorithmically. Designed for test-retest reliability.

Modal participle

State layer

Session-specific mode. Sensitive to current signals. Lower retest reliability by design.

15 participles × 14 nouns = 210+ unique descriptor pairs computed from quantitative profile. Not selected from a fixed library.

Example · Computed output
Stabilizing · Coordinator
Noun source BFI-44 profile
Participle source Session signals
Retest (noun) High by design
Retest (participle) Variable by design
STRUCTURAL NOUN Coordinator deterministic · derived from profile dimensions Dimension source BFI-44 projection Test-retest designed for stability maps to: trait dimension (Spielberger et al., 1970) MODAL PARTICIPLE Stabilizing session-specific · generated from behavioral signals Dimension source divergence signals Test-retest state-sensitive · expected to vary maps to: state dimension (Spielberger et al., 1970) COMPUTED OUTPUT: "Stabilizing Coordinator"
Figure 3 Two-component archetype decomposition. The structural noun is derived deterministically and carries high test-retest reliability. The modal participle is state-sensitive and is expected to vary across sessions — by architectural design, not as a limitation.
05 · Divergence Patterns

The three divergence axes.

The Shift layer is operationalized as a function of measurable disagreement between observation channels. Each divergence axis provides a distinct observation invisible from any single channel.

Channels What the divergence reveals Example pattern
Self-report ↔ Speech Discrepancy between how the user characterizes themselves and the statistical structure of their unconstrained verbal output under rhythmic pressure. High-Conscientiousness self-report; speech pattern shows repeated topic switching and unfinished ideas — neither contradicts the other.
Speech ↔ Facial Discrepancy between the affective valence of verbal content and the concurrent facial expression signals captured during the same session. Neutral or composed verbal tone; facial signals indicate elevated arousal or suppressed emotional leakage invisible in the verbal channel.
Self-report ↔ Facial Discrepancy between the user's described emotional state and the nonverbal signals present during the session — a second independent axis to localize which channel is anomalous. Self-reported Agreeableness; facial microexpression analysis reveals tension patterns inconsistent with the self-description.
SAMPLE DIVERGENCE PROFILE Self-report Verbal Facial Conscientiousness Agreeableness Openness Neuroticism Extraversion Divergence Index 0.41 0.18 0.74 ▲ 0.33 HIGH DIVERGENCE → Shift layer content Self-report Verbal output Facial signal Divergence index (0–1)
Figure 4 Sample divergence profile across five personality dimensions. Three channels shown as independent rows. The divergence index (rightmost column) quantifies inter-channel disagreement per dimension. High-divergence dimensions (≥0.5) generate Shift layer content.
06 · Publication

Submitted for review.
Publicly archived.

Preprint · PsyArXiv · May 2026 · Pending peer review

Recognition + Shift

A Two-Layer Framework for AI-Mediated Behavioral Self-Reflection
Authors D. Stolyarov · S. Stolyarov · I. Stolyarov
Version v0.5 · 10 May 2026
Repository PsyArXiv · OSF Preprints
License CC-BY 4.0 International
Subject area Social and Behavioral Sciences
Patent status 2 patents pending (RU, 2026) · PCT Oct 2026
Abstract (condensed)

The growing class of AI-generated behavioral reports faces a recognized paradox: outputs users readily confirm as accurate often fail to produce insight. We name this the plausible-but-flat problem and propose Recognition + Shift — a two-layer design principle in which the second layer is operationalized strictly through cross-channel divergence, not interpretation.

4 Novelty Claims
  • Named and formalized the plausible-but-flat problem
  • Cross-channel divergence as primary source of behavioral insight
  • Two-component state-trait archetype with empirically testable retest pattern
  • Three-channel minimum requirement for interpretable divergence
07 · References

The foundations.
Not the claim.

FaceMirror did not invent triangulation, state-trait distinction, or multimodal observation. These are established ideas. What we built is a specific architecture that applies them together in a way that has not been applied before. The sources below are the shoulders we stand on — not the claim we make.

Personality theory Triangulation methodology State-trait architecture Behavioral observation +Computational implementation ↓
Personality theory Triangulation State-trait Behavioral observation + Computational · HCI · Affective computing
01
The Fallacy of Personal Validation
Forer, B. R. · J. Abnormal & Social Psychology · 1949

Foundation of the plausible-but-flat problem. Barnum effect in personality assessment.

02
Wanted — A Good Cookbook
Meehl, P. E. · American Psychologist · 1956

Justification for the deterministic structural noun over LLM interpretation.

03
Convergent and Discriminant Validation by the Multitrait-Multimethod Matrix
Campbell & Fiske · Psychological Bulletin · 1959

Direct methodological foundation of the three-channel triangulation architecture.

04
STAI Manual for the State-Trait Anxiety Inventory
Spielberger et al. · Consulting Psychologists Press · 1970

Precedent for the two-component archetype: structural noun (trait) + modal participle (state).

05
The Research Act
Denzin, N. K. · McGraw-Hill · 1978

Formalization of triangulation: divergence between methods as information, not error.

06
The Big Five Inventory
John & Srivastava · Handbook of Personality · 1999

Validation of the 44-item instrument used as Channel 1 of the triangulation system.

07
Facial Action Coding System (FACS)
Ekman & Friesen · Consulting Psychologists Press · 1978

Foundation of the facial expression channel (Channel 3) and its emotional signal framework.

08
Personality in Adulthood
McCrae & Costa · Guilford Press · 2003

Five-factor model stability; context for trait-level versus state-level observation.

09
Word Association and Language Processing
Deese, J. · Johns Hopkins University Press · 1965

Empirical basis for free verbal association as a behavioral observation channel.

10
Lie to Me: A Field Guide
Ekman, P. · W. W. Norton · 2009

Cross-channel divergence between verbal content and facial expression as meaningful signal.

11
Personality Psychology: Domains of Knowledge
Larsen & Buss · McGraw-Hill · 2008

Overview of multi-method assessment; theoretical context for the triangulation approach.

12
The Handbook of Emotional Intelligence
Bar-On & Parker · Jossey-Bass · 2000

Context for behavioral self-reflection as a distinct category from trait assessment.

C1
Affective Computing
Picard, R. W. · MIT Press · 1997

Foundational framework for computational recognition of affect from multiple behavioral signals — direct precursor to multimodal channel fusion.

C2
Honest Signals: How They Shape Our World
Pentland, A. · MIT Press · 2008

Multimodal behavioral signal analysis: nonverbal channels carry structural information not accessible from self-report or verbal content alone.

C3
Social Signal Processing
Vinciarelli et al. · IEEE Signal Processing Magazine · 2009

Computational framework for automatic analysis of social signals from audio, visual, and combined channels — computational legitimacy for the three-channel architecture.

08 · Scope & Limits

What this framework
does not claim.

Scientific honesty requires explicit limits. The following boundaries are built into the framework specification, not added as legal disclaimers after the fact.

Not clinical diagnosis

FaceMirror outputs are not psychiatric or psychological diagnoses. The framework explicitly avoids clinical categorization. Results should not be used as substitutes for professional clinical assessment.

Not forensic or legal use

Results are not valid for employment screening, legal proceedings, forensic evaluation, or any context requiring certified psychometric evidence. These uses are explicitly prohibited in the framework specification.

Not unconscious content access

The metronome technique shifts the balance from edited self-presentation toward less-filtered associative output — modestly, at the margin. It does not claim unmediated access to unconscious processes.

Not perfectly reproducible

The same input does not produce identical output across language-model runs. The structural noun replicates at high reliability; the modal participle does not — by design. This is a feature of the architecture, not a defect.

Not a substitute for psychometrics

FaceMirror does not substitute for established psychometric instruments. Its claim is design-level and phenomenological — producing a qualitatively different user experience than single-layer reports.

What it is

A multimodal self-reflection tool for behavioral awareness. Designed to surface observable cross-channel patterns invisible from any single observation channel. For personal development and professional self-understanding only.