Production Architecture
M.I.N.D. v3.0 Pipeline
Content enters as a signal. Each layer scores an orthogonal manipulation dimension. The layer scores combine into a single Manipulation Potential Score (MPS).
L1
Pattern FrequencyKeyword density, repetition anomalies, timing patterns
Production
L2
Emotional ArcEngineered sentiment trajectories, emotional manipulation
Production
L3
Network PropagationCascade fingerprints, coordinated amplification
Production
L4
Choice ArchitectureFalse dichotomies, anchoring, manipulative option framing
Production
L5
Behavioral LoopCommitment escalation, reciprocity exploitation, habit cycles
Production
L6
Cognitive LoadInformation pacing targeting critical thinking capacity
Production
Detection Modules - 68 Production
| Directorate | Modules | Focus |
|---|---|---|
| CogArc | 20 | Cognitive bias exploitation, strategic framing, memetic analysis |
| Experimental | 22 | State influence ops, radicalization pipelines, dark personality detection |
| CogArc v3 | 10 | Overton shifts, staircase escalation, psychographic targeting |
| Experimental v3 | 10 | Algorithmic amplification, multimodal dissonance, code-switching |
| Sentinel | 6 | AI-agent defense: prompt injection, jailbreak, alignment probing |
Formal Languages
- HPL v3.0: Custom DSL for detection rules. Variables, functions, weighted scoring, boolean composition, cross-directorate chaining.
- APL: AI-to-human manipulation vectors - prompt injection signatures, persona manipulation, agent coordination.
Integration Surfaces
- Python SDK (
mind-py) and JavaScript SDK (mind-js) with full parity - Telegram bot (30+ commands) and Discord bot (full parity)
- Browser extension (Chrome + Firefox) with cognitive firewall overlay
- REST API (v3 transport layer) and Docker deployment
Internal Validation
Benchmark Results
| Benchmark | N | Precision | FPR | Recall | Notes |
|---|---|---|---|---|---|
| Curated | 165 | 1.000 | 0.000 | 0.986 | Hand-labeled by domain experts. F1=0.993. |
| Synthetic | 50,000 | 1.000 | 0.000 | 0.393 | Template-based, deterministic (seed=42). Internal only. |
Qualifications:
Curated benchmark is small (165 samples, 25 benign). Results should be interpreted with appropriate uncertainty.
Synthetic benchmark is template-generated. Performance may not generalize to real-world content.
No independent third-party replication has been conducted.
Precision-first design accepts ~39% recall on synthetic data as the cost of zero observed false positives.
Curated benchmark is small (165 samples, 25 benign). Results should be interpreted with appropriate uncertainty.
Synthetic benchmark is template-generated. Performance may not generalize to real-world content.
No independent third-party replication has been conducted.
Precision-first design accepts ~39% recall on synthetic data as the cost of zero observed false positives.
Proof-of-Concept Research Extensions
v9.0 Research Roadmap
Layers 7–14 and additional directorates are in proof-of-concept stage. None contribute to production MPS scores. All evaluation on internal synthetic data only.
L7
Discourse TrajectoryMulti-turn escalation, grooming sequences
Partial
L8
Identity PalaceLongitudinal persona tracking
Concept
L9
Psychological ResonanceDeep psychological pattern analysis
Concept
L10
Cross-User CorrelationCross-user exposure correlation
Concept
L11
Recursive MonitoringAdversary adaptation, feedback cycles
Concept
L12
Collective DynamicsEmergent belief structures, cascade dynamics
Concept
L13
Epistemic AuthorityNarrative control detection
Concept
L14
Active InferencePredictive processing exploitation detection
PoC
Not production technology. Layer 14 uses phrase-based heuristics (not variational inference). 15 modules in the Predictive Noetic Directorate. Default weight 0.0. Preliminary synthetic signal only (~+1.2 pp recall, experimental profile). Requires external validation on naturalistic corpora.
Contact
General: contact@esothel.com · Research: research@esothel.com · Security: security@esothel.com
Get in Touch