Profiles
Governance profiles
Profiles define how X-40™ evaluates a workload class and what policy it enforces. Use a profile per use case (facts, math, unknowns, attack, or ML inference governance).
facts
- Short factual questions and business answers.
- Stability + drift control; verification routing when evidence is insufficient.
- Designed to reduce “confident but wrong” shipments in everyday factual flows.
math
- Arithmetic and deterministic computation prompts.
- Deterministic math guard: accept only if verified.
- Designed to minimize Wrong+Accepted risk in numeric workflows.
unknowns
- Unanswerable or fictitious queries.
- Enforces unknown/refusal handling (e.g., returns UNKNOWN instead of fabrication).
- Best for legal, compliance, and regulated knowledge workflows.
attack
- Prompt injection and malicious instruction patterns.
- Forbidden-output detection + safe refusal enforcement.
- Best for enterprise copilots and internal assistants exposed to hostile inputs.
ML profiles (examples)
ML integrations govern inference using confidence telemetry and drift signals. Profiles define thresholds, escalation actions, and retention posture.
ml_finance
Strict drift thresholds and audit-first posture for high-stakes decisions.
ml_risk
Conservative accept; rapid verification routing when drift rises.
ml_legal_review
High verification bias for sensitive categories; privacy-max defaults.
Example ML telemetry payload (conceptual):
{
"profile": "ml_risk",
"prediction": "approve",
"confidence": 0.93,
"margin": 0.21,
"batch_id": "2025-12-28T09",
"trace_mode": "none"
}