Overview

/manipulation-detect is a production command in the Defensive Security category of the Prismatic Platform that performs active detection of manipulation attempts against the platform's epistemic systems using multi-layered analysis techniques. The command scans intelligence streams, agent behavior patterns, confidence distributions, and knowledge base integrity to identify attempts to distort truth, manipulate confidence, poison signals, induce drift, or hijack salience within the platform's decision-making infrastructure.

This command operates under the L3 authority level and is executed by the manipulation-detector agent. It is part of the platform's 216-command slash command registry, built on the AIAD (Autonomous Intelligence Agent Design) standard. The L3 authority level reflects the sensitive nature of manipulation detection -- the command requires access to agent behavior data, intelligence streams, and confidence scoring internals to perform effective detection.

Manipulation detection in the Prismatic Platform is grounded in the NABLA epistemic framework's seven non-negotiable axioms. Each axiom defines a property that must hold for the platform's knowledge to be trustworthy: signal plurality, contradiction preservation, absence informativeness, time decay, unknown validity, source independence, and provenance mandatoriness. The manipulation detection system monitors for violations of these axioms as indicators of manipulation. A sudden reduction in signal plurality may indicate source poisoning. Disappearing contradictions may indicate truth distortion. Confidence scores clustering around specific values may indicate confidence manipulation.

The detection engine uses five primary analysis techniques that correspond to the five categories of epistemic attack identified by the Red Team: fact verification (detects truth distortion), statistical analysis (detects confidence manipulation), source validation (detects signal poisoning), baseline comparison (detects drift induction), and attention analysis (detects salience hijacking). These techniques operate continuously in the background and can also be invoked on-demand for specific scope scanning.

Architecture

The detection system operates as a multi-technique analysis engine with continuous monitoring and on-demand scanning capabilities.

Detection Architecture

/manipulation-detect -> Scan Coordinator -> Analysis Engines -> Verdict Engine
                              |                   |                   |
                              v                   v                   v
                        Scope Definition    Fact Verifier       Threat Score
                        Target Selection    Stats Analyzer      Classification
                        Baseline Loading    Source Validator     Evidence Chain
                        Signal Collection   Drift Detector      Remediation
                                            Salience Analyzer

Detection Techniques

Technique	Target Attack	Method	Indicators
Fact Verification	Truth Distortion	Compare assertions against knowledge base	Contradiction rate, source deviation
Statistical Analysis	Confidence Manipulation	Analyze confidence score distributions	Bimodal clustering, artificial plateaus
Source Validation	Signal Poisoning	Verify signal provenance and plurality	Single-source dominance, provenance gaps
Baseline Comparison	Drift Induction	Compare current behavior against established baselines	Deviation velocity, cumulative drift
Attention Analysis	Salience Hijacking	Analyze attention distribution across targets	Focus concentration, topic deviation

Detection Scope Levels

Scope	Targets	Duration	Use Case
Agent	Single agent behavior and outputs	1-5 minutes	Targeted investigation
Channel	Intelligence stream or data feed	5-15 minutes	Stream integrity check
Domain	Analysis domain (financial, legal, etc.)	10-30 minutes	Domain-wide assessment
Pipeline	Complete intelligence pipeline	15-60 minutes	Full pipeline audit
Platform	All platform epistemic systems	30-120 minutes	Comprehensive scan

Confidence Scoring for Detections

Confidence	Threshold	Meaning	Action
Confirmed	> 0.95	Multiple independent indicators confirm manipulation	Immediate defense activation
Probable	0.75 - 0.95	Strong indicators across multiple techniques	Alert + investigation
Possible	0.50 - 0.75	Single technique indicates manipulation	Monitor + investigate
Unlikely	0.25 - 0.50	Weak indicators, likely false positive	Log + monitor
Not Detected	< 0.25	No meaningful indicators	Normal operations

Usage

# Run platform-wide manipulation scan
/manipulation-detect --scope=platform

# Scan specific agent
/manipulation-detect --scope=agent --target=sig-osint-commander

# Scan intelligence channel
/manipulation-detect --scope=channel --target=osint-pipeline

# Scan with specific technique focus
/manipulation-detect --technique=truth-distortion --scope=domain --target=financial

# Continuous monitoring mode
/manipulation-detect --continuous --interval=60

# Quick triage scan
/manipulation-detect --scope=platform --depth=quick

# Deep analysis scan
/manipulation-detect --scope=agent --target=ma-financial-analyst --depth=exhaustive

# Scan with custom baseline
/manipulation-detect --scope=agent --target=sig-osint-commander \
  --baseline="2026-01-01:2026-02-01"

# Export detection results
/manipulation-detect --scope=platform --format=json --output=detection-results.json

# Run Red Team scenario validation
/manipulation-detect --validate-scenario=red-team-scenario-042

# Scan specific time window
/manipulation-detect --scope=platform --time-window="2026-02-14T00:00:00Z/2026-02-15T00:00:00Z"

# Detection with sensitivity adjustment
/manipulation-detect --scope=platform --sensitivity=high

Options

Option	Type	Default	Description
`--scope`	string	platform	Detection scope: agent, channel, domain, pipeline, platform
`--target`	string	all	Target agent, channel, or domain name
`--technique`	string	all	Detection technique: truth-distortion, confidence-manipulation, signal-poisoning, drift-induction, salience-hijacking, all
`--depth`	string	standard	Scan depth: quick, standard, deep, exhaustive
`--continuous`	flag	false	Enable continuous monitoring
`--interval`	integer	300	Continuous monitoring interval in seconds
`--baseline`	string	auto	Baseline period for comparison (start:end dates)
`--sensitivity`	string	normal	Detection sensitivity: low, normal, high, maximum
`--format`	string	text	Output format: text, json, markdown
`--output`	string	stdout	Output file path
`--validate-scenario`	string	none	Red Team scenario ID for validation
`--time-window`	string	last-24h	Time window for analysis (ISO 8601 interval)
`--confidence-threshold`	float	0.5	Minimum confidence for reported detections
`--include-false-positives`	flag	false	Include suspected false positives in output

Execution Flow

Scope Resolution: The detection scope is resolved to a specific set of monitoring targets. Agent scope resolves to a single agent's behavior data. Channel scope resolves to a specific intelligence stream. Platform scope includes all monitoring points.
Baseline Loading: Behavioral baselines for all targets in scope are loaded from the baseline database. Baselines include normal confidence distributions, typical signal patterns, established fact references, and historical attention patterns.
Signal Collection: Current signals are collected from all targets in scope. This includes agent outputs, intelligence stream data, confidence scores, knowledge base state, and attention metrics for the configured time window.
Parallel Analysis: All five detection techniques execute in parallel against the collected signals. Each technique compares current observations against baselines and detection rules to identify anomalies.
Anomaly Scoring: Each detected anomaly is scored for manipulation confidence. Scoring considers anomaly magnitude, consistency across multiple indicators, duration, and correlation with known attack patterns.
Cross-Technique Correlation: Anomalies detected by individual techniques are correlated to identify coordinated attacks. A truth distortion anomaly combined with a confidence manipulation anomaly on the same target significantly increases the overall manipulation confidence score.
Classification: Confirmed detections are classified against the manipulation technique taxonomy. The classification includes the specific technique variant, estimated sophistication level, affected scope, and potential impact.
Evidence Chain Construction: For each detection, an evidence chain is constructed linking the detected anomaly to the specific signals, baseline deviations, and analysis results that support the detection. This chain provides full auditability.
Alert Generation: Detections exceeding the confidence threshold generate alerts that are sent to /manipulation-dashboard and, if configured, to /manipulation-protect for automated defense activation.
Result Reporting: Detection results are formatted and output according to the configured format and destination.

Integration Points

Component	Relationship	Details
Prismatic Agents	Executed by `manipulation-detector`	Detection analysis engine
/manipulation-protect	Defense activation	Detections trigger protection responses
/manipulation-dashboard	Visualization	Detection signals feed dashboard
/manipulation-techniques	Classification	Taxonomy for detection classification
Color-Team	Red/Blue framework	Red Team validates, Blue Team defends
NABLA Framework	Axiom monitoring	Detection based on NABLA axiom violations
Quality Gates	Quality integration	Detection quality validation
Telemetry	Event tracking	Detection events logged for analysis
/emergency	Crisis escalation	Critical detections trigger emergency protocol

Best Practices

Run continuous monitoring during intelligence operations. Active OSINT collection sessions introduce external data that may contain manipulation attempts. Continuous monitoring during these sessions provides real-time detection coverage.

Use appropriate sensitivity for the threat environment. Normal sensitivity is suitable for routine operations. Increase to high during active threat periods or when processing intelligence from untrusted sources.

Validate with Red Team scenarios. Periodically run --validate-scenario against Red Team attack simulations to verify that the detection system identifies known attack patterns. Detection gaps against known attacks indicate tuning needs.

Review false positives periodically. Use --include-false-positives to review near-threshold detections that were classified as unlikely. Patterns in false positives may indicate emerging attack techniques that current rules underweight.

Establish baselines during calm periods. Baselines established during periods of known-clean operation are more reliable than those established during uncertain periods. Use --baseline to specify clean reference periods.

Investigate drift detections thoroughly. Drift induction is specifically designed to be sub-threshold -- each individual change is small enough to avoid detection. When drift is detected, the cumulative effect is often significant and warrants thorough investigation.

Error Handling

Error	Cause	Resolution
`target_not_found`	Specified agent or channel does not exist	Verify target name with system inventory
`baseline_not_available`	No baseline data for specified period	Establish baseline first or use auto-baseline
`signal_collection_timeout`	Could not collect signals in time	Reduce scope or increase timeout
`technique_not_available`	Specified technique module not loaded	Check module availability, restart if needed
`scenario_not_found`	Red Team scenario ID does not exist	Verify scenario ID with Red Team inventory
`insufficient_data`	Not enough data for meaningful analysis	Extend time window or wait for more data

Advanced Usage

Custom Detection Rules

Add organization-specific detection rules.

# Add custom confidence manipulation rule
/manipulation-detect --add-rule --name="custom-confidence-check" \
  --technique=confidence-manipulation \
  --condition="confidence_stdev < 0.05 AND sample_size > 100" \
  --severity=high

# Test custom rule against historical data
/manipulation-detect --test-rule="custom-confidence-check" --time-window="2026-01-01/2026-02-01"

Forensic Analysis Mode

Deep analysis of suspected manipulation events.

# Forensic analysis of specific event
/manipulation-detect --forensic --event-id=MAN-2026-042 --depth=exhaustive

# Trace manipulation chain across agents
/manipulation-detect --trace --start-event=MAN-2026-042 --follow-propagation

Integration with External Threat Intelligence

Incorporate external threat indicators into detection.

# Import threat indicators
/manipulation-detect --import-indicators=external-threats.json

# Cross-reference with known attack patterns
/manipulation-detect --scope=platform --cross-reference=threat-db

Doctrine Compliance

All commands operate under the NO MERCY, NO DOUBTS doctrine:

NO MERCY: Zero tolerance for incomplete execution or quality violations. Detection scans complete fully or report exactly which components could not be scanned. No monitoring gap is silently accepted.
NO DOUBTS: Full investigation before action, evidence-based results. Every detection includes a complete evidence chain, confidence scoring, and classification against the established taxonomy. Detections without evidence are not reported.

/manipulation-protect - Activate manipulation protection defenses
/manipulation-techniques - View manipulation technique taxonomy and counter-measures
/manipulation-dashboard - Manipulation detection dashboard with threat indicators
/emergency - Emergency response and crisis management activation
/archer-supreme - Supreme authority activation for platform-wide operations
/dark-ops - NABLA structural crisis detection and dark operations analysis

Connect & Contribute

Created by Tomáš Korcak (korczis) | Open Source under GHL

GitHub | GitLab | LinkedIn | Contact
Developer Portal | Architecture | Meet the Creator

/manipulation-detect