I II III IV

Esothel Labs

Cognitive Security · Influence Defense · Adversarial Intelligence

Private applied-research laboratory. We build defensive systems that detect manipulation in language - not after the damage is done, but while it’s happening.

Founded 2025. Lean team. No venture theater. Real technology shipping to real environments.

Influence-Operation Defense · Adversarial Pattern Detection · Cognitive Resilience Infrastructure

// About

What We Do

Information warfare is the defining threat of this decade. State actors run influence operations across 80+ countries. AI-generated persuasion content is scaling faster than human analysts can review it. Legacy content-moderation systems are optimized for virality metrics, not cognitive defense.

We build the other side of that equation.

Esothel Labs develops detection systems that identify manipulation at the psychological mechanism level - not fact-checking, not sentiment analysis, not keyword blacklists. We detect how content is engineered to exploit cognitive vulnerabilities: manufactured urgency, precision-targeted framing, identity exploitation, belief-cascade engineering, and coordinated inauthentic behavior patterns.

Everything we build is defensive. We do not create influence operations. We detect, classify, and neutralize them.

// Current Production Capability

M.I.N.D. - Modeling Influence, Neutralizing Deception

M.I.N.D. v3.0 is our core detection framework, currently deployed and operational.

Pipeline6 weighted analysis layers, each targeting an orthogonal manipulation dimension
Modules68 detection modules across core directorates (CogArc, Experimental, Sentinel)
DetectionPhrase/rule-based heuristics + lightweight ML priors + structural pattern analysis
OutputManipulation Potential Score (MPS) on a 0–1 continuous scale with discrete threat levels
LanguagesHPL v3.0 (Human Persuasion Language) - custom DSL for formal threat specification
PlatformsPython SDK, JavaScript SDK, Telegram bot, Discord bot, browser extension, REST API

The Six Production Layers

LayerNameWhat It Detects
L1Pattern FrequencyUnnatural repetition, keyword density anomalies, timing patterns
L2Emotional ArcEngineered emotional trajectories, sentiment manipulation sequences
L3Network PropagationCoordinated amplification signatures, cascade fingerprints
L4Choice ArchitectureFalse dichotomies, manufactured scarcity, manipulative framing
L5Behavioral LoopCommitment escalation, intermittent reinforcement, habit engineering
L6Cognitive LoadInformation pacing designed to overwhelm critical thinking

Validation

  • Internal benchmark: ~205 samples (165 threat + 40 benign), hand-curated by domain experts
  • Results on that set: Precision 1.000, FPR 0.000, zero false positives observed
Important qualifier: These results are on a small, internally curated benchmark. Independent third-party validation on larger naturalistic corpora is pending and is a priority. We do not claim these metrics generalize to all real-world content.
Deployments
Telegram & Discord bots  ·  Chrome/Firefox browser extensions  ·  REST + WebSocket API  ·  Python & JavaScript SDKs  ·  Docker single-command deploy

Design Philosophy

Precision over recall. M.I.N.D. is calibrated to never flag legitimate content as manipulative. We accept lower recall - missing some manipulation - as the explicit cost of maintaining operator trust. When M.I.N.D. flags something, you can act on it. That’s the point.

// Research Pipeline

Where We’re Heading

Beyond the production core, Esothel Labs maintains an active research program exploring next-generation detection approaches grounded in computational neuroscience.

Active Inference Research Extensions (v9.0)

We are developing a theoretical framework that maps manipulation detection onto the Free Energy Principle (Friston, 2010) and predictive processing theory. The core insight: manipulative content can be characterized as adversarial perturbation of the brain’s generative models.

  • Layer 14 (Active Inference) implemented as a proof-of-concept research layer
  • 15 Predictive Noetic Directorate modules targeting specific predictive-processing exploitation mechanisms
  • Default weight: 0.0 - Layer 14 does not contribute to production MPS scores
  • Detection: phrase-based heuristics only (not deep semantic or variational inference computation)
  • Evaluation: internal synthetic benchmark only, deterministic (seed=42), no external validation
  • Preliminary internal signals suggest complementary detection on synthetic data (~+1.2 pp recall)
This is not production technology. It is a serious research direction with real theoretical grounding (Friston, Hohwy, Seth, Clark, Rao & Ballard) and promising early signals. We discuss it openly because we believe in transparency about where our technology is heading - but we do not sell what we haven’t validated.

Research Roadmap

  • Near-term: External validation on naturalistic corpora. Independent replication.
  • Medium-term: Embedding-based semantic detection (beyond phrase matching). Public benchmark contributions.
  • Long-term: Physiological validation studies (EEG, HRV, MMN) connecting text-based detection to measurable cognitive effects.

// Publications

Research Output

14 manuscripts in preparation for arXiv submission (with full qualification of metrics and limitations).

Foundation

Core Framework & Architecture Papers (2)

Six-layer pipeline architecture, internal benchmark results, configurable detection modules.

Evaluation

Scaling & Recall Studies (2)

Preliminary evaluation on large-scale synthetic benchmarks. Recall hardening methodology with auto-tuned HPL rules.

Theory

Detection Framework Papers (4)

Epistemic authority detection, collective belief dynamics, adversary adaptation modeling, coordinated influence detection. All theoretical/proof-of-concept.

Theory

Active Inference Series (6)

Theoretical frameworks mapping manipulation detection to free energy minimization, predictive coding, collective phase transitions. Proof-of-concept implementations.

Disclaimer: All manuscripts describe research/proof-of-concept work. Production claims are limited to M.I.N.D. v3.0 core (6 layers, 68 modules). All evaluation beyond the ~205-sample curated benchmark is on internal synthetic data only. No independent external validation has been conducted on research extensions. Full publication list with honest summaries available upon request.

// Applications

Who This Is For

  • Defense and intelligence: Cognitive domain defense, influence campaign attribution, population resilience assessment
  • Government and policy: Information environment monitoring, election integrity, counter-disinformation infrastructure
  • Enterprise security: Employee cognitive security training, executive protection, supply-chain information integrity
  • Research institutions: Academic collaboration on active inference, predictive processing, physiological validation
  • Venture studios and investors: Due diligence on cognitive security as an emerging category

Contact

Serious inquiries from government, defense, enterprise, and research partners.

General: contact@esothel.com  ·  Research: research@esothel.com  ·  Security: security@esothel.com

We respond to substantive inquiries within 48 hours. No pitch decks on request - we demo live capability.

Get in Touch