Skip to content

Gödel Guardrails Scoring Cloud / Team

Automated quality scoring for Standard Operating Procedures (SOPs) using a 13-category rubric that gates SOP lifecycle transitions.

What Is the Gödel Score?

The Gödel score is a quality metric on a 0–26 scale that evaluates whether an SOP is comprehensive, consistent, and safe enough to be enforced by the governance engine. Named after Kurt Gödel's work on formal system completeness, it ensures that SOPs meet a minimum quality bar before they can influence agent behavior.

Every SOP submitted to the control plane is scored against 13 rubric categories, each worth 0–2 points. The total score determines which lifecycle gate the SOP passes through.


Threshold Gates

The Gödel score controls automatic SOP lifecycle transitions:

Score RangeGate ResultEffect
< 12BLOCKEDSOP cannot proceed — it is rejected and must be revised
12–14PENDING_REVIEWSOP requires manual review by an Admin or EM before advancing
≥ 15GENERATEDSOP auto-advances to the next lifecycle state

TIP

These thresholds can be tuned via environment variables to match your organization's risk tolerance. See Configuration below.

Lifecycle Integration

The Gödel score integrates into the SOP lifecycle at the DRAFT → PENDING_REVIEW / GENERATED transition:

DRAFT → [Gödel Scoring] → BLOCKED (score < 12)
                         → PENDING_REVIEW (12 ≤ score < 15)
                         → GENERATED (score ≥ 15)

SOPs that score BLOCKED remain in DRAFT state and surface a quality report showing which rubric categories scored low.


The 13-Category Rubric

Each category is scored from 0 (not addressed) to 2 (fully addressed):

#CategoryWhat It Evaluates
1CompletenessDoes the SOP cover all necessary steps and edge cases?
2ConsistencyAre instructions internally consistent and non-contradictory?
3TestabilityCan compliance with the SOP be objectively verified?
4Side-Effect CoverageAre potential side effects of agent actions documented?
5Error HandlingDoes the SOP specify what to do when things go wrong?
6SecurityAre security implications addressed (secrets, permissions, data exposure)?
7Scope ClarityIs it clear when this SOP applies and when it does not?
8Version CompatibilityDoes the SOP account for different tool/API versions?
9Performance ImpactAre resource usage and performance considerations documented?
10ObservabilityDoes the SOP include logging, metrics, or alerting guidance?
11MaintainabilityIs the SOP structured for easy updates as requirements evolve?
12DocumentationIs the SOP itself well-documented with examples and rationale?
13Adversarial ResilienceDoes the SOP resist prompt injection or manipulation attempts?

Scoring Example

A well-written SOP covering code review requirements might score:

Completeness:            2/2
Consistency:             2/2
Testability:             2/2
Side-Effect Coverage:    1/2  (missing rollback steps)
Error Handling:          2/2
Security:                2/2
Scope Clarity:           2/2
Version Compatibility:   1/2  (no version-specific notes)
Performance Impact:      2/2
Observability:           2/2
Maintainability:         2/2
Documentation:           1/2  (examples could be richer)
Adversarial Resilience:  1/2  (no injection guard clauses)
─────────────────────────────
Total:                  22/26 → GENERATED ✅

Configuration

Tune the gate thresholds with environment variables:

VariableDefaultDescription
GODEL_SCORE_THRESHOLD_GENERATED22Minimum score for automatic advancement to GENERATED
GODEL_SCORE_THRESHOLD_BLOCKED12Scores below this are BLOCKED

WARNING

Lowering GODEL_SCORE_THRESHOLD_GENERATED reduces the quality bar for auto-approved SOPs. Only adjust this if you have a robust manual review process for PENDING_REVIEW SOPs.

Scores between GODEL_SCORE_THRESHOLD_BLOCKED and GODEL_SCORE_THRESHOLD_GENERATED land in PENDING_REVIEW, requiring manual approval from an Admin or EM.


Current Implementation

Phase 1 — Static Scoring

The current implementation (Phase 1) returns a static score of 22 for all SOPs, placing them directly in the GENERATED gate. This ensures forward compatibility while the full LLM-based scoring engine is developed in Phase 2.

Phase 2 will integrate real 13-category LLM-based scoring via the Gödel Guardrails evaluator service, where each SOP's markdown content is evaluated against the rubric by an LLM judge.


The active firewall for AI agents