Govern AI across every framework.

KoraSafe builds governance artifacts from first principles, then maps each field to the regulations that apply to your organization. One workflow, any framework.

Governance principles

What every framework agrees on

Strip away the regulatory labeling and a consistent set of requirements emerges across EU AI Act, GDPR, ISO 42001, NIST AI RMF, and SR 11-7. KoraSafe builds its governance model around those shared requirements, not around any single regulator's terminology.

Risk classification

Every framework tiers AI systems by the potential harm of a failure. KoraSafe's risk register captures the inputs that determine tier, independent of which framework is doing the scoring.

Human oversight for consequential decisions

Consequential AI decisions require a human in the loop who can review, override, and be accountable. KoraSafe captures oversight design in the HITL Spec artifact and validates it at runtime.

Data governance and bias monitoring

Training data lineage, representation audits, and subgroup performance checks are required by every major framework. KoraSafe's Data Provenance and Bias Testing editors capture the evidence.

Continuous monitoring and incident response

Post-deployment drift, performance degradation, and incident tracking are not optional add-ons. KoraSafe's Post-Market Monitoring Plan and guardian runtime agents keep evidence current.

Documentation as evidence

Governance claims without artifacts are assertions. KoraSafe generates structured, auditable documentation for every stage of the AI lifecycle, timestamped and hash-verified.

Transparency

Dual transparency model

Regulators and affected persons need different transparency artifacts. KoraSafe generates both from a single data source, keeping them in sync as the model evolves.

Model Card

Technical transparency for developers, auditors, and downstream deployers. Documents intended use, performance benchmarks, training data, limitations, and known failure modes.

  • Intended use and out-of-scope uses
  • Subgroup performance across protected attributes
  • Training data sources and known gaps
  • Evaluation methodology and benchmarks
  • Recommendations for downstream deployers

System Card

Operational transparency for compliance teams, regulators, and senior stakeholders. Documents how the AI is deployed, what decisions it influences, and what human oversight is in place.

  • System architecture and data flows
  • Consequential decision scope and affected persons
  • Human oversight design and accountability chain
  • Incident response and escalation paths
  • Post-market monitoring commitments
AI value chain

Where you sit in the chain

Different frameworks use different vocabulary for supply chain roles, but the concepts are consistent. Knowing your role determines which obligations apply.

Role 1

Provider

Develops or places an AI system on the market. Bears primary documentation and conformity obligations.

Role 2

Deployer

Uses an AI system in a professional context. Responsible for use-case risk assessment and human oversight.

Role 3

Importer

Places a third-country AI system on a regulated market. Verifies provider compliance before deployment.

Role 4

Distributor

Makes an AI system available without modifying it. Checks compliance and cooperates with authorities.

An organization can hold multiple roles simultaneously. KoraSafe's system registry captures role assignments per AI system and applies the corresponding obligation set.

Residual risk

Risk that stays after controls

Residual risk documentation captures what remains after mitigations are applied. It is a pre-deployment artifact required by ISO 42001 Section 6.1, NIST AI RMF Manage 2.1, EU AI Act Article 9, and SR 11-7 model risk governance. KoraSafe generates it from the Risk Register and requires board-level sign-off before the system enters production.

What goes into the Residual Risk artifact

Inherent risk level, controls applied and their effectiveness, residual risk rating, accepted residual risk rationale, board approval record, and scheduled review cadence. Each field maps to the frameworks that require it.

ISO 42001 §6.1 NIST AI RMF Manage 2.1 EU AI Act Art. 9 SR 11-7 model risk
Conformity

What conformity actually requires

Across the frameworks KoraSafe supports, conformity for high-risk AI systems converges on the same seven workstreams. KoraSafe generates artifacts for each.

01

Technical documentation

Structured record of system design, data, training, evaluation, and intended use. KoraSafe generates the outline from the 9-section universal template.

02

Risk management

Continuous risk identification, assessment, mitigation, and residual risk acceptance across the full system lifecycle.

03

Quality management system

Documented policies, procedures, and controls for development, deployment, and change management. Required by ISO 42001 and EU AI Act Article 17.

04

Human oversight design

Defined override mechanisms, accountability chains, and review cadence for consequential decisions. Captured in the HITL Spec and validated at runtime.

05

Model Card and System Card

Transparency artifacts for developers and compliance stakeholders respectively. Both generated from a single governance data source.

06

Logging and audit trail

Automated, tamper-evident event logging for all consequential decisions. KoraSafe's hash-chained audit trail satisfies logging requirements across frameworks.

07

Post-market monitoring

Structured plan for tracking performance, incidents, and emerging risks after deployment. KoraSafe generates the plan and keeps it current via runtime guardian data.

Explainability

Explainability under human oversight

A human-in-the-loop cannot meaningfully override a decision they cannot understand. Explainability is therefore a prerequisite for effective human oversight, not a separate requirement. KoraSafe supports the tooling that bridges the two.

Tabular models

SHAP

SHAP (SHapley Additive exPlanations) decomposes a model's output into per-feature contributions. KoraSafe's SHAP starter produces subgroup-aggregated outputs in the bias testing import format.

Non-tree models

LIME

LIME (Local Interpretable Model-agnostic Explanations) fits a simple surrogate model in the neighborhood of each prediction. Useful for image and text classifiers where SHAP is computationally heavy.

LLM decisions

Counterfactuals

DiCE counterfactual explanations surface the minimal input changes that would flip a decision. Particularly relevant for adverse action notices under FCRA and automated decision explanations under GDPR Article 22.

Starter scripts for all three approaches are available in the developer documentation.

Regulatory mapping

Frameworks KoraSafe maps to

KoraSafe maps your governance work to the frameworks that apply to your organization. The mapping is metadata on each artifact field, surfaced as framework coverage chips. Your governance model stays neutral. The mapping updates as frameworks evolve.

Framework Jurisdiction / sector Key obligations KoraSafe covers
EU AI Act EU,all sectors Technical doc, risk management, HITL, logging, FRIA, post-market monitoring, conformity declaration
GDPR EU,personal data processing DPIA, Art. 22 automated decision rights, data minimization, lawful basis documentation
ISO 42001 International,all sectors AI management system, risk treatment, QMS, supplier governance, continual improvement
NIST AI RMF US,all sectors Govern, Map, Measure, Manage functions; risk tolerance documentation; AI risk register
SR 11-7 US,banking Model risk management, validation, ongoing monitoring, model inventory, governance policies
NYC Local Law 144 US,NYC employment Automated employment decision tool bias audit, public results disclosure, candidate notice
Colorado SB 205 US,Colorado, consequential decisions Algorithmic discrimination assessment, impact assessment, adverse action notice
FCRA US,consumer credit Adverse action notice, principal reason codes, explainability for credit decisions
HIPAA US,healthcare PHI handling in AI pipelines, de-identification validation, minimum necessary principle
CCPA / CPRA US,California consumers Automated decision-making opt-out, data provenance, purpose limitation

Per-framework deep dives are in progress. Contact us if you need a specific framework covered sooner.

Scoring methods

How KoraSafe measures its claims.

Scoring models, statistical methods, and probe design, documented so your audit team can verify the numbers.