Conversational Jev (Simple Jev) bridges free-form natural language intent with deterministic System 1 decision-making models. Because user text flows into an LLM generation pipeline that constructs executable schemas, security and determinism are architectural priorities.

If you discover a security vulnerability within Conversational Jev, please do not open a public issue.

Instead, report it via responsible disclosure:

1. Send an email to the project maintainers or open a GitHub Private Vulnerability Report.
2. Include a description of the issue, steps to reproduce, and any proof-of-concept payload or curl command.
3. We will acknowledge receipt within 48 hours and provide an estimated resolution timeline.

Free-form user input is parsed by a Generator LLM to produce candidate schemas. Without controls, adversarial inputs could attempt to hijack the LLM prompt. Conversational Jev implements defense-in-depth:

• Pre-Sanitization Barrier: backend/app/services/security.py scans inbound text for prompt-injection signatures, delimiters (e.g., <system>, [INST], markdown backtick escapes), and role-reversal patterns.
• Strict Output Confinement: The Generator LLM produces structured JSON schemas according to rigid Pydantic models (CandidateSchema). Free-form text generation outside of the schema definition is discarded.
• Untrusted Author Isolation: The Generator LLM is explicitly treated as an untrusted author. Its output is never executed directly.

Even if a prompt-injection payload manipulates the Generator LLM into proposing an unusual schema, the 5-Point Meta-Schema Validator audits the candidate schema prior to any execution:

• Mutual Exclusivity: Verifies options do not overlap or conflict.
• Scope & Sizing: Disallows arbitrary option proliferation or unbounded categories.
• State Sufficiency: Confirms that state parameters contain sufficient factual grounding before execution can proceed.
• Type Safety: Strictly validates against Jev's primitive definitions (Choice, Score, Noul).

The intent cache utilizes a 384-dimensional dense vector index (all-MiniLM-L6-v2) to provide sub-millisecond execution for approved intents:

• No Cross-Tenant Data Leaks: Cache entries index normalized semantic intent and approved decision templates. Raw sensitive user data, PII, and credentials are removed prior to vector indexing.
• Failure Memory: The cache tracks rejected or stalled schemas to prevent recurring degenerate loops.
• Divergence Guard: Any semantic mutation or novelty automatically forces a pause for explicit user confirmation in Restricted Mode.

The Jev decision engine is deterministic and non-autoregressive. Unlike generative models that can hallucinate or leak context during generation, Jev evaluates typed question schemas against closed state dictionaries. It cannot run arbitrary code, spawn external network calls, or read local file systems.

• All API keys (JEV_API_KEY, GEMINI_API_KEY, GROQ_API_KEY, OPENAI_API_KEY) are read strictly from server-side environment variables via backend/app/config.py.
• No credentials or API keys are bundled into frontend assets or transmitted over client-side WebSocket/HTTP payloads.
• In development and demonstration environments, the system defaults to deterministic simulation mode when keys are omitted.