Skip to main content
Jira progress: loading…

AI-DRDT

AI Decision Registry & Codex Decision Tables

1. Overview

ZAYAZ makes extensive use of AI-assisted workflows to enrich, classify, validate, and normalize information entering the platform. These workflows may operate on imported spreadsheets, registry records, telemetry, documents, APIs, or other structured and semi-structured data.

Rather than allowing AI-generated outputs to modify production registries directly, every AI decision is recorded and made reviewable before promotion into the platform's canonical registries.

This provides:

  • Complete auditability
  • Explainable AI decisions
  • Human review and approval
  • Reproducible processing
  • Provider-independent AI governance

The AI decision architecture consists of two layers:

  1. Legacy Codex workflow tables
  2. Canonical AI Decision Registry

The Codex tables continue to support existing enrichment workflows and remain part of the platform architecture, while the AI Decision Registry provides a generalized, future-proof mechanism for recording all AI-generated decisions regardless of the underlying model or provider.


2. Architectural Overview

            Source Data


AI / LLM Processing

┌────────┴─────────┐
│ │
▼ ▼
Legacy Codex Tables AI Decision Registry
│ │
└────────┬─────────┘

Human Review


Canonical ZAYAZ Registries

The platform deliberately separates:

  • operational AI processing
  • AI decision evidence
  • production business data

This separation is essential for governance, explainability, and regulatory compliance.


3. Codex Tables

The existing codex_* tables support specific AI-assisted workflows used during registry generation, enrichment, validation, and classification.

Although originally named after the early Codex concept, these tables are not tied to any particular AI product or provider. They may be populated by OpenAI models, Anthropic models, local models, AWS Bedrock, Azure AI, or other future AI systems.

codex_automation_runs

Records every AI-assisted workflow execution.

Typical information includes:

  • workflow type
  • execution status
  • execution timing
  • orchestration metadata
  • input source
  • processing statistics
  • operational outcome

This table provides operational traceability for AI execution.


codex_signal_registry_working

Working table used while constructing or enriching the Signal Registry.

Signals placed here are considered candidates and may undergo:

  • normalization
  • deduplication
  • enrichment
  • classification
  • validation
  • human review

Only approved signals are promoted into the canonical Signal Registry.


codex_signal_name_results

Stores AI-generated signal naming recommendations.

Typical outputs include:

  • normalized signal names
  • suggested canonical names
  • confidence scores
  • rationale
  • review metadata

The objective is to prevent duplicate or inconsistent signal naming across the platform.


codex_datatype_results

Stores AI-generated datatype classifications.

Examples include:

  • Text
  • Integer
  • Decimal
  • Boolean
  • Date
  • Timestamp
  • JSON

This assists schema generation, validation, and registry enrichment.


codex_kind_results

Stores semantic classifications describing what a record represents.

Examples include:

  • Signal
  • Metric
  • Organization
  • Certification
  • Registry Entity
  • Component
  • Module

Unlike datatype classification, Kind describes the semantic meaning of an object.


codex_module_results

Stores AI-generated mappings between records and the canonical ZAYAZ module hierarchy.

Examples include:

  • Environmental
  • Social
  • Governance
  • Input Hub
  • Reports Hub
  • Computation Hub

This supports automated routing and registry organization.


codex_validator_results

Stores AI-assisted validation results.

Validation may include:

  • rule checking
  • unit validation
  • semantic validation
  • completeness checks
  • consistency checks
  • governance rules

Validation outputs include:

  • pass/fail status
  • confidence
  • rationale
  • suggested remediation

4. AI Decision Registry

While the Codex tables support individual workflows, the AI Decision Registry provides a generalized governance layer capable of storing every AI-generated decision made within the platform.

This architecture is provider-independent and designed for long-term scalability.


ai_decision_runs

ai_decision_runs represents the parent execution record for an AI workflow.

Each row represents one complete execution.

Typical information includes:

  • workflow type
  • execution environment
  • model provider
  • model version
  • input source
  • processing statistics
  • execution status
  • timestamps
  • orchestration metadata

One execution may generate hundreds or thousands of individual AI decisions.

Example:

Run 9845

Import:
ESRS Metrics Workbook

Records processed:
12,483

Model:
GPT-5.5

Status:
Completed

ai_decision_registry

ai_decision_registry is the canonical evidence store for AI-generated decisions.

Each row represents one decision produced during an AI workflow.

Examples include:

  • datatype classification
  • module classification
  • signal naming
  • ontology mapping
  • validator result
  • registry enrichment
  • semantic classification
  • recommendation
  • confidence assessment

Each decision records:

  • decision type
  • decision value
  • confidence score
  • rationale
  • model metadata
  • prompt version
  • review status
  • reviewer
  • timestamps

The registry provides complete explainability and supports future regulatory requirements around AI governance.


5. Relationship Between the Tables

The relationship is intentionally hierarchical.

ai_decision_runs (1)

├──────────────┐
│ │
▼ ▼
Decision A Decision B
│ │
▼ ▼
ai_decision_registry (many)

One AI workflow may generate thousands of individual decisions, all linked back to the execution that produced them.

This design provides complete lineage from:

  • source data
  • AI execution
  • individual decisions
  • human review
  • final registry promotion

6. esign Principles

The AI decision architecture follows several core principles.

Explainability

Every AI-generated conclusion should be traceable and explainable.

Auditability

AI decisions are preserved even if they are later rejected or superseded.

Human Oversight

Critical decisions may require review before affecting production registries.

Provider Independence

The architecture is independent of any specific AI provider and supports future migration between models and vendors.

Reproducibility

Recording model versions, prompts, confidence scores, and execution metadata allows AI-assisted workflows to be reproduced and audited over time.


7. Summary

The Codex tables provide specialized support for existing AI-assisted registry workflows.

The AI Decision Registry provides a generalized governance framework capable of recording every AI-generated decision across the ZAYAZ platform.

Together they establish a scalable foundation for explainable AI, operational traceability, regulatory compliance, and future AI governance while ensuring that no AI-generated information is promoted into production without a complete evidence trail.




GitHub RepoRequest for Change (RFC)