See the stars

ContextKernel Specification

Home / ContextKernel / ContextKernel Specification

ContextKernel

Measured Signals for Serious Moments.


Overview

ContextKernel is an open-source, privacy-preserving artificial intelligence specification designed to identify potential indicators of intent to harm others through contextual analysis of conversational interactions. Rather than relying on isolated keywords or individual statements, ContextKernel uses large language models to evaluate the meaning, progression, and circumstances of questions and conversations. It is designed to recognize patterns such as escalating hostility, targeted fixation, expressed threats, potential preparation, and other behavioral indicators that may warrant further assessment.

The system is designed to operate within encrypted communication environments without weakening existing encryption, introducing centralized plaintext message collection, or creating unauthorized access to private conversations. Its architecture prioritizes local processing, data minimization, cryptographic accountability, and transparent risk assessment.

ContextKernel evaluates conversational patterns over time and generates narrowly scoped notifications when predefined risk thresholds warrant human review. Notifications can include specific text excerpts, relevant timestamps, contextual explanations, confidence assessments, and supporting evidence for authorized monitors. Automated classifications are treated as indicators requiring human assessment rather than determinations of guilt, dangerousness, or future behavior.

The specification incorporates modular capabilities for contextual language analysis, behavioral trajectory assessment, intent classification, LLM training, privacy-preserving inference, encrypted communication integration, selective disclosure, and human-reviewed intervention. Its extensible architecture supports additional research, evaluation, oversight, and notification capabilities while maintaining transparency, individual privacy, and accountability.


Core Modules

Contextual Language Analysis Module

  • Analyze questions, statements, and conversational exchanges for contextual meaning.
  • Evaluate individual messages in relation to surrounding conversational context.
  • Distinguish hypothetical discussions, academic research, fictional scenarios, emotional expressions, and genuine indications of harmful intent.
  • Identify language suggesting targeted hostility, threats, coercion, intimidation, or expressed intent to harm.
  • Recognize indirect, ambiguous, coded, and evolving expressions that may require additional context.
  • Distinguish discussions about violence from apparent advocacy, preparation, or intention to commit violence.
  • Identify uncertainty and insufficient evidence rather than forcing definitive classifications.
  • Support contextual analysis across different communication styles and conversational formats.

Behavioral Trajectory Analysis Module

  • Evaluate changes in conversational patterns over time.
  • Identify escalating patterns of hostility, fixation, grievance, and expressed intent.
  • Recognize repeated discussions involving particular individuals, groups, organizations, or locations.
  • Analyze transitions from generalized frustration to targeted threats or apparent preparation.
  • Identify combinations of behavioral indicators that may warrant further human assessment.
  • Distinguish isolated statements from persistent patterns.
  • Support configurable temporal analysis windows.
  • Account for contradictory statements and changes in conversational context.
  • Avoid interpreting personality characteristics, political beliefs, religion, ethnicity, or other protected characteristics as indicators of dangerousness.

Intent Classification Module

  • Classify potentially harmful intent using contextual evidence.
  • Distinguish emotional distress, violent ideation, threatening language, expressed intent, and apparent preparation.
  • Identify references to potential victims, stated objectives, timelines, and conditional threats.
  • Evaluate whether statements describe past events, hypothetical situations, fictional scenarios, or prospective actions.
  • Assess the specificity and apparent credibility of expressed threats.
  • Identify uncertainty, contradictory statements, and missing contextual information.
  • Provide explainable classifications with supporting evidence.
  • Avoid assigning permanent dangerousness labels to individuals.
  • Maintain distinctions between concerning language, credible threats, and potentially imminent danger.

LLM Training and Fine-Tuning Module

  • Train and fine-tune language models for contextual safety and behavioral risk classification.
  • Support supervised fine-tuning and parameter-efficient model adaptation.
  • Incorporate publicly available research, appropriately licensed datasets, expert-reviewed examples, and synthetic scenarios.
  • Separate training examples involving fictional, hypothetical, historical, and real-world threatening situations.
  • Support adversarial training against ambiguous language and misleading conversational context.
  • Evaluate model consistency across different communication styles and languages.
  • Implement dataset provenance tracking and licensing verification.
  • Support differential privacy techniques where appropriate.
  • Maintain reproducible training configurations and model evaluation records.
  • Publish model cards, limitations, and documented training methodologies.
  • Prevent private user conversations from being silently incorporated into training datasets.

Risk Assessment and Confidence Module

  • Combine contextual, behavioral, and temporal indicators into an explainable assessment.
  • Maintain separate classifications for severity, immediacy, specificity, and confidence.
  • Establish configurable thresholds for human review.
  • Distinguish potential threats from credible, specific, and apparently imminent threats.
  • Account for conflicting evidence and contextual uncertainty.
  • Apply conservative escalation policies to reduce false positives.
  • Record the evidence supporting each classification.
  • Provide confidence estimates and explanations of model limitations.
  • Support independent evaluation of classification accuracy.
  • Prevent automated risk scores from independently triggering punitive actions.

Privacy-Preserving Processing Module

  • Perform language analysis locally wherever supported.
  • Keep decrypted message content within the authorized endpoint or existing trusted processing boundary.
  • Preserve the encryption architecture of integrated communication platforms.
  • Avoid creating centralized plaintext conversation repositories.
  • Minimize the information retained during analysis.
  • Provide configurable retention and deletion policies.
  • Support ephemeral processing and secure memory handling.
  • Prevent unauthorized access to message content and analysis results.
  • Ensure model training does not silently incorporate private conversations.
  • Document all information processed, retained, and disclosed.
  • Make data processing and disclosure behavior independently auditable.

Encrypted Communication Integration Module

  • Provide integration interfaces for encrypted messaging platforms.
  • Support compatible communication clients and decentralized messaging environments.
  • Integrate with locally available message content without introducing server-side decryption.
  • Preserve existing end-to-end encryption and key-management arrangements.
  • Support client-side deployment and trusted execution environments where available.
  • Provide configurable integration permissions.
  • Prevent integrations from accessing conversations outside their explicitly authorized scope.
  • Support independently reviewable integration adapters.
  • Provide compatibility interfaces for open communication protocols.
  • Clearly document the security boundaries and permissions of each integration.

Alert Generation and Notification Module

  • Generate alerts when defined risk thresholds warrant human review.
  • Include specific supporting text excerpts rather than unrestricted conversation exports.
  • Provide contextual explanations for flagged material.
  • Identify the relevant classification, confidence, and assessment limitations.
  • Include timestamps and narrowly scoped conversational context where authorized.
  • Support configurable alert recipients and escalation policies.
  • Encrypt notifications and supporting evidence in transit and at rest.
  • Prevent unauthorized recipients from accessing sensitive excerpts.
  • Support alert acknowledgment, review, dismissal, and escalation.
  • Maintain records of alert generation and subsequent human decisions.
  • Provide configurable notification categories based on severity and urgency.

Selective Disclosure Module

  • Generate narrowly scoped evidence packages for authorized reviewers.
  • Separate risk notifications from access to underlying message content.
  • Support cryptographic authorization before releasing protected excerpts.
  • Implement multi-party approval requirements where configured.
  • Provide configurable disclosure limits and expiration periods.
  • Maintain cryptographic hashes to verify the integrity of disclosed evidence.
  • Prevent access to unrelated conversational material.
  • Record the scope, authorization, recipient, and time of each disclosure.
  • Support revocation of access to future retrievals where technically possible.
  • Ensure that disclosure policies are transparent and independently auditable.

Human Review and Intervention Module

  • Provide authorized monitors with contextual risk assessments.
  • Present the evidence supporting each alert.
  • Distinguish model-generated conclusions from verified facts.
  • Allow reviewers to dismiss, downgrade, escalate, or request additional context.
  • Require human assessment before consequential interventions.
  • Provide procedures for urgent, credible threats involving immediate danger.
  • Support documented intervention recommendations.
  • Record reviewer decisions and explanations.
  • Provide mechanisms for correcting inaccurate classifications.
  • Prevent automated enforcement based solely on model-generated risk scores.
  • Support independent review of disputed classifications.

Cryptographic Accountability Module

  • Sign model releases, configurations, and assessment components.
  • Verify the integrity of deployed models.
  • Maintain tamper-evident audit records.
  • Support cryptographic hashing of selected evidence.
  • Implement threshold cryptography for multi-party disclosure authorization.
  • Protect notification channels and reviewer access.
  • Provide cryptographic verification of evidence packages.
  • Detect unauthorized modifications to system configurations.
  • Maintain verifiable records of disclosure and access events.
  • Support independent verification of system integrity.

Model Evaluation and Red-Team Module

  • Evaluate classification accuracy across different conversational contexts.
  • Measure false-positive and false-negative rates.
  • Test the system against fictional scenarios, academic discussions, satire, and ambiguous statements.
  • Evaluate performance across languages and communication styles.
  • Test adversarial prompting and contextual manipulation.
  • Identify systematic bias and disproportionate classification patterns.
  • Evaluate the reliability of confidence estimates.
  • Measure performance under incomplete conversational context.
  • Conduct regression testing for every model release.
  • Publish independent evaluation results and documented limitations.
  • Support external research and reproducible benchmarking.

Transparency and Audit Module

  • Provide public documentation of model behavior and limitations.
  • Publish risk classification definitions and assessment criteria.
  • Document training datasets and their provenance.
  • Maintain versioned model evaluation reports.
  • Support independent security and privacy audits.
  • Provide reproducible builds and signed releases.
  • Document every category of information collected, processed, or disclosed.
  • Provide transparent alert and escalation policies.
  • Maintain public records of significant model and policy changes.
  • Support transparency reporting for alert generation and disclosure activity.
  • Document security vulnerabilities and corrective measures.

Optional Plugin Modules

Additional Language Support Plugin

  • Support multilingual conversational analysis.
  • Detect contextual expressions and threats across supported languages.
  • Account for regional linguistic differences and idiomatic expressions.
  • Evaluate translation uncertainty before risk classification.
  • Provide language-specific evaluation datasets.
  • Support independent validation of multilingual classification accuracy.

Imminent Threat Assessment Plugin

  • Evaluate indicators of potentially immediate physical danger.
  • Assess combinations of expressed intent, specificity, timing, and apparent preparation.
  • Identify statements suggesting an immediate threat to identifiable individuals or groups.
  • Support configurable urgent human-review procedures.
  • Distinguish immediate threats from generalized or historical expressions.
  • Provide supporting evidence and uncertainty assessments.

Multimodal Context Plugin

  • Support analysis of authorized images, audio transcripts, and other conversational attachments.
  • Evaluate relationships between text and associated media.
  • Maintain separate privacy permissions for each media category.
  • Prevent unauthorized extraction or retention of multimedia content.
  • Support contextual analysis of authorized multimedia material.
  • Document the limitations of multimodal classification.

Decentralized Monitoring Plugin

  • Support distributed alert delivery across independently operated monitoring organizations.
  • Implement authenticated notification endpoints.
  • Support decentralized authorization and verification.
  • Prevent centralized accumulation of conversational data.
  • Provide configurable distribution policies.
  • Maintain verifiable records of notification delivery and access.

Research and Dataset Development Plugin

  • Provide tools for creating synthetic training scenarios.
  • Support expert-reviewed behavioral classification datasets.
  • Enable independent academic evaluation.
  • Track dataset licensing, provenance, and annotation quality.
  • Support reproducible benchmark generation.
  • Provide tools for identifying dataset bias and annotation inconsistencies.

Multilingual Threat Research Plugin

  • Support research into culturally and linguistically dependent expressions of threats.
  • Evaluate model performance across languages.
  • Identify translation-related classification errors.
  • Support language-specific validation datasets.
  • Evaluate contextual differences in idioms and communication patterns.

Independent Oversight Plugin

  • Provide independent review of classification policies.
  • Support privacy and security compliance assessments.
  • Document false-positive investigations.
  • Provide transparency reports and audit summaries.
  • Support external governance and oversight organizations.
  • Review changes to risk thresholds and disclosure policies.

Incident Documentation Plugin

  • Organize authorized evidence packages.
  • Maintain incident timelines and reviewer decisions.
  • Support cryptographically verifiable documentation.
  • Provide controlled evidence export for authorized investigations.
  • Track review status and documented outcomes.
  • Enforce configurable evidence retention policies.

Model Comparison Plugin

  • Compare multiple language models against standardized evaluation datasets.
  • Measure differences in classification accuracy and confidence calibration.
  • Identify inconsistent results across models.
  • Evaluate performance under adversarial and ambiguous conditions.
  • Support independent model selection and validation.
  • Maintain historical evaluation records across model versions.

Privacy Verification Plugin

  • Evaluate whether deployed components comply with documented privacy requirements.
  • Verify that message processing remains within authorized security boundaries.
  • Audit data retention and disclosure configurations.
  • Detect unauthorized changes to privacy settings.
  • Verify model signatures and deployment configurations.
  • Generate independently reviewable privacy assessment reports.

Specification Branding License (SBL)

Standard

Optional


License & Notice Requirements

ContextKernel is released under the GNU Affero General Public License v3.0 or later (AGPL-3.0+).
By contributing to any Open Arsenal project, you agree that your contributions will also be released under this license.

Please note the following:

  • All contributions must comply with the AGPL-3.0+ terms.
  • Under Section 7 of the license, all redistributions, forks, and derivative works must preserve attribution to:
    Roxanne Ardary and roxanneardary.com.
  • ContextKernel specifications are free to use with attribution. A Specification Branding License can be negotiated upon request.
  • The project’s notice.md file tracks attribution requirements and contributor acknowledgments.
    Any update that adds new contributors or modifies attribution should also update notice.md.
  • When submitting a pull request, ensure that any new files maintain the attribution headers where applicable.
  • Network-deployed versions of this software must also remain fully AGPL-3.0+ compliant, including exposure of source code modifications when applicable under the license.

For full legal details, please refer to the AGPL-3.0+ license and the project’s notice.md file.


Notice – ContextKernel

Attribution Requirement: Under Section 7 of the AGPL 3.0+ license, all redistributions, forks, and derivative works, including network-deployed versions of this project, must provide attribution to Roxanne Ardary and roxanneardary.com.

Contributors

This file tracks contributors and their specific contributions to the project.

  • Roxanne Ardary, roxanneardary.com – November 18, 2026
    Created the repository for ContextKernel. Established the project concept, modular architecture, privacy-preserving behavioral risk detection framework, LLM training specifications, cryptographic accountability requirements, and human-reviewed notification system.
  • [Add other contributors here] – [Date]
    [Describe contribution in one sentence]

License – ContextKernel

This repository is licensed under the GNU Affero General Public License v3.0 or later (AGPL-3.0+).

Key Points:

  • You are free to use, modify, and distribute the code.
  • All redistributions, forks, and derivative works or network-deployed versions must also be licensed under AGPL-3.0+ and provide attribution to Roxanne Ardary and roxanneardary.com as required under Section 7 of the license.
  • The software is provided “as is,” without warranty of any kind.

For the full license text, see GNU AGPL-3.0 License.