Adversarial Psychology / Pattern Cognition

PAREIDOLIA

Coherence Injection as Attack Surface
The Signal That Was Never There

Every human analyst is a pattern-recognition engine with no off switch. PAREIDOLIA maps the threshold between noise and committed false belief, and asks what happens when that threshold is found, held, and exploited.

Active Research Behavioral / Adversarial
01 The Condition

In 1958, German psychiatrist Klaus Conrad coined apophenia to describe a symptom he observed in schizophrenic patients: the unmotivated experience of meaningful connections between unrelated things. He named the earliest stage Trema -- a state of charged, anxious readiness -- and the second stage Apophanie, from the Greek apo (away) and phaenein (to show): the moment hidden meaning breaks through the surface of the world.

Conrad framed this as pathology. He was half right. What he described is not a malfunction -- it is the human pattern-recognition system running at full capacity in an environment that does not warrant it. The system is not broken. It is being misled.

Pareidolia is the perceptual subspecies: seeing faces in clouds, figures in static, intent in noise. It is apophenia with a sensory mask. The brain's object-recognition circuits fire before the frontal lobe can audit the input, producing conviction before evidence.

The evolutionary logic is unambiguous. In a world where a shadow might be a predator, the cost of a false positive is embarrassment. The cost of a false negative is death. Natural selection built a system biased toward detection, not accuracy. The brain does not ask "is there a pattern?" It asks "what is the most plausible pattern given this data?" -- and then commits.

This bias is not historical. It is running right now, in every intelligence analyst studying satellite imagery, every security researcher reading logs, every radiologist reading a scan. The same circuitry that saw faces in cave walls sees intent in noise, clusters in scatter, trends in drift.

PAREIDOLIA begins with a single question: if the bias cannot be turned off, can it be aimed?

02 Signal Detection Theory and the Bias Parameter

Signal Detection Theory, formalized by Green and Swets in 1966, separates two quantities the human mind conflates: sensitivity (d-prime) -- the true ability to distinguish signal from noise -- and criterion (beta) -- the internal threshold at which the observer commits to "yes, there is something here."

Sensitivity is relatively stable across context. Criterion is not. It shifts under cognitive load, emotional arousal, sleep deprivation, and social pressure. An analyst told that "something is likely here" will lower their criterion without realizing it, generating more false positives while believing they are being appropriately vigilant. The signal they find feels real because it was found, not because it is there.

Hit
TP
Signal present, analyst responds. The expected outcome -- weaponized as a baseline trust that makes false positives plausible.
False Alarm
FP
Signal absent, analyst responds. The target state. Maximized by lowering criterion without degrading the analyst's perceived sensitivity.
Miss
FN
Signal present, analyst does not respond. Too low a false alarm rate indicates the injected coherence is insufficient.
Correct Rejection
TN
Signal absent, analyst does not respond. The attack's failure state. Coherence is below the commitment threshold.

The adversarial insight is this: an attacker does not need to overcome an analyst's sensitivity. They need only to shift their criterion. A primed analyst will commit to a false pattern that an unprimed analyst would ignore -- not because they are less capable, but because their internal threshold has been relocated without their awareness.

The Receiver Operating Characteristic curve maps the full space of possible tradeoffs. A skilled analyst operates upper-left: high hit rate, low false alarm rate. The PAREIDOLIA attack does not push the analyst off the curve. It moves them along it, toward a region of higher false alarm rate that they experience as appropriate vigilance.

03 Predictive Processing and the Prior

Karl Friston's predictive processing framework reframes perception entirely. The brain does not passively receive sensory input and interpret it. It generates predictions about the world -- hypotheses -- and compares them to incoming data. What we experience as perception is the brain's best guess, corrected by prediction error.

This architecture has a critical property: the prior matters. A confident expectation suppresses prediction error. The brain sees what it expects to see and discounts contradictory evidence. Friston calls this precision-weighted prediction error minimization. A high-precision prior is difficult to dislodge not because of logic, but because of architecture.

The adversarial implication is direct: if you can install a prior, you can shape perception -- not by changing the data, but by changing what the analyst expects to find in it.

Intelligence briefings, threat assessments, and peer statements are all prior-installers. "We believe there may be activity in sector 7" is not a neutral statement. It loads a prior that will resist disconfirmation. Subsequent ambiguous data resolves in favor of the prior because that is how the perceptual system is built -- not as a failure of judgment, but as an expression of Bayesian inference operating on manipulated inputs.

The two-stage PAREIDOLIA model exploits this directly. Stage one is prior installation: credible, low-commitment signals that establish the expectation of a pattern. Stage two is coherence injection: structured noise calibrated to confirm what the analyst already half-believes.

The prior does most of the work. The injected coherence only needs to be sufficient to confirm an expectation, not to produce one from nothing.

04 The Attack Model
Prior Installation
Phase 01 / Criterion Shift
Credible context is established before data is introduced. Briefings, authority signals, peer consensus, or anomaly reports lower the analyst's detection criterion without presenting false data. Only framing changes. The analyst is not deceived yet -- they are prepared to deceive themselves.
Coherence Injection
Phase 02 / Signal Calibration
Structured noise is introduced into the data stream. Coherence is calibrated to the individual threshold: high enough to trigger commitment, low enough to appear organic. The analyst finds the pattern they were primed to find and experiences it as discovery, not as a reaction to planted structure.
Commitment Lock
Phase 03 / Belief Crystallization
Once the analyst commits verbally or in writing, confirmation bias and sunk-cost dynamics maintain the belief. Contradictory data is processed as anomalous rather than disconfirming. The false pattern becomes self-sustaining without further input from the attacker.
Threshold Variance
Key Variable / Individual Calibration
Detection threshold varies significantly between individuals and across contexts. Fatigue, cognitive load, domain expertise, and emotional state all shift criterion. A well-calibrated attack targets the distribution, not a single threshold value, accepting some false negatives in exchange for undetectable coherence levels.

The critical design constraint is the detectability boundary. Too little coherence and the analyst correctly rejects the noise. Too much and the fabrication becomes visible -- the structure is too clean, the pattern too convenient. The attack operates in the narrow band between these limits, where coherence is sufficient to produce commitment but insufficient to produce suspicion.

This band is not fixed. It shifts with context, expertise, and cognitive state. The instrument below is designed to locate it.

05 Threshold Calibration -- Live Instrument

The instrument below presents a live signal stream. Each trial begins as pure noise. Coherence is injected at a randomized, hidden rate. Press PATTERN DETECTED the moment you commit to seeing structure in the waveform. Your threshold -- the coherence level at the moment of commitment -- is recorded. Five trials build your response distribution.

You are the subject.

06 Thesis
Research Position

Adversarial research has focused almost entirely on technical surfaces: code, networks, protocols, models. The human analyst is treated as the trusted endpoint -- the final verification layer that cannot be compromised without direct deception.

This framing is wrong. The human analyst is not the verification layer. They are the most reliably exploitable surface in the chain -- not through deception, but through the fundamental architecture of perception itself. Apophenia is not a vulnerability in the analyst. It is a vulnerability in the species.

PAREIDOLIA does not require lying. It requires understanding how the brain resolves ambiguity -- and providing the right kind of ambiguity at the right moment. The pattern that commits an analyst is not false. It is simply not there. The difference is invisible from the inside.

Understanding the threshold is the first step toward defending it. The instrument is the map.

FIELD ASSESSMENT

The instrument above measures your threshold in a neutral context. The field assessment runs the full attack pipeline -- prior installation, coherence injection, commitment -- without disclosure. You will not be told what is happening until it is over.

ENTER FIELD ASSESSMENT