Skip to main content

​
Purpose

Risk scores in TealTiger provide a deterministic, quantitative signal that represents how risky a request, action, or execution context is across security, cost, and reliability. Rather than relying solely on binary allow/deny outcomes, TealTiger evaluates risk on a normalized spectrum, enabling:
  • Graduated enforcement
  • Context-aware decisions
  • Explainable governance
  • Policy-driven automation
Risk scores are a foundational input used by policies, reason codes, and decision resolution.

​
Risk Score Contract (Normative)

This section is normative. Any risk score produced by TealTiger MUST conform to the following rules.

​
Required properties

A risk score MUST include:
  • score — a normalized numeric value in the range 0–100
  • level — a categorical interpretation of the score
  • version — the risk model version used to compute the score

​
Optional properties

A risk score MAY include:
  • domain_scores — per-domain breakdowns for security, cost, and reliability
  • signals — identifiers of contributing signals (when exposed)
Applications should not rely on raw scoring formulas. They should rely on score, level, and policy thresholds.

​
Risk Score Range and Scale

TealTiger uses a normalized numeric range:
  • 0 — no meaningful risk
  • 100 — maximum unacceptable risk
The range is intentionally fixed to ensure:
  • comparability across domains
  • stable policy thresholds
  • consistent audit interpretation

​
Risk Levels (Categorical)

Numeric scores are mapped to categorical risk levels for human and policy readability. The following levels are defined:
  • LOW — acceptable risk
  • MEDIUM — elevated risk; review or restriction may apply
  • HIGH — unacceptable risk; enforcement likely required
Exact numeric boundaries between levels are implementation details. Policies should target levels or ranges, not internal formulas.

​
Domain-Specific Risk

Risk in TealTiger is intentionally multi-dimensional.

​
Security risk

Security risk reflects the likelihood or impact of:
  • unauthorized tool access
  • data exfiltration or leakage
  • policy circumvention attempts

​
Cost risk

Cost risk reflects the likelihood or impact of:
  • excessive token usage
  • budget overruns
  • inefficient model or tool selection

​
Reliability risk

Reliability risk reflects the likelihood or impact of:
  • timeouts and retries
  • cascading failures
  • degraded service quality
Domain risks may be combined into a composite score or evaluated independently by policies.

​
Composite vs Domain Scores

  • score represents the composite risk.
  • domain_scores (when present) represent per-domain contributors.
Policies may:
  • act on the composite score
  • act on individual domain scores
  • combine both approaches
The combination logic is deterministic and versioned.

​
How Risk Scores Are Used

Risk scores influence decision-making in several ways:
  • policy thresholds (e.g., block if score > X)
  • graduated actions (allow, transform, degrade, deny)
  • reason code emission (explaining which threshold was crossed)
  • enforcement mode behavior (monitor vs enforce)
Risk scores do not directly execute actions. They are signals consumed by policy evaluation.

​
Determinism and Stability Guarantees

TealTiger guarantees:
  • Risk scoring is deterministic for the same inputs and model version.
  • Risk scores are versioned via risk.version.
  • Meanings of score ranges and levels are stable within a version line.
  • New signals or refinements may be added only in additive, non-breaking ways.

​
What Risk Scores Are Not

Risk scores are intentionally limited:
  • ❌ They are not probabilistic predictions
  • ❌ They do not learn or adapt at runtime
  • ❌ They do not override policy logic
  • ❌ They are not free-form explanations
Risk scores are signals, not decisions.

​
Example Risk Score (Conceptual)

​
Relationship to Decisions and Audit Events

  • Risk scores inform policy evaluation.
  • Decisions determine actions.
  • Reason codes explain why thresholds were crossed.
  • Audit events capture the risk score and decision as evidence.

​
Evolution Rules

  • ✅ Add new risk signals (additive)
  • ✅ Add new domain-specific scores
  • ❌ Change the meaning of existing score ranges
  • ❌ Change the normalized score range
  • ❌ Introduce non-deterministic scoring
  • /concepts/decision-model
  • /concepts/decision-lifecycle
  • /policy/reason-codes
  • /policy/conditions-and-actions
  • /audit/audit-event-schema