Skip to main content
Open Source Python Package

ISAF Logger

Evidence Generation for AI Training Lineage

Add 3 lines of code to get cryptographic lineage evidence for AI training pipelines. Logs Layers 6-9 (Framework, Data, Objectives, Deployment) with SHA-256 hash chains. Generates evidence that can support compliance documentation. Works with PyTorch, TensorFlow, JAX, and scikit-learn.

View on GitHubQuick Start
$pip install isaf-logger

Why ISAF Logger?

AI regulations require documentation of how your AI systems were trained. ISAF generates lineage evidence — framework versions, data sources, objective functions, and deployment configuration — with cryptographic hash chains that prove the evidence has not been tampered with.

3 Lines of Code

Add compliance logging to any ML training pipeline with minimal code changes. Works with your existing workflow.

Full Stack Coverage

Automatically logs Layer 6 (Framework), Layer 7 (Data), and Layer 8 (Objectives) of the instruction stack.

Cryptographic Verification

SHA-256 hash chains prove lineage integrity. Tamper-evident audit trails that support compliance documentation.

Compliance Mapping References

Maps logged data to EU AI Act Article 10 & 11, NIST AI RMF, ISO 42001, and Colorado AI Act requirements. These are references to support compliance documentation - not certification.

Framework Agnostic

Works with PyTorch, TensorFlow, JAX, and scikit-learn. Auto-detects your ML framework.

Flexible Storage

SQLite for local development, MLflow integration for production. Export to JSON anytime.

Quick Start

Get compliance logging in under 5 minutes

1. Install

$ pip install isaf-logger

2. Add to Your Code

import isaf

# Initialize ISAF (one line)
isaf.init()

# Add decorators to your training functions
@isaf.log_data(source="customer_data", version="3.2.1")
def load_training_data():
    return pd.read_csv("data.csv")

@isaf.log_objective(
    name="binary_crossentropy",
    constraints=["fairness < 0.05"]
)
def train_model(data):
    model = create_model()
    model.fit(data)
    return model

# Run training as normal
data = load_training_data()
model = train_model(data)

# Export compliance report (one line)
isaf.export("compliance_report.json")

What Gets Logged

6

Layer 6: ML Framework

Framework versions, CUDA availability, default parameters, numerical precision

7

Layer 7: Training Data

Data source, version, shape, dtypes, missing values, preprocessing operations

8

Layer 8: Objective Function

Loss function, mathematical form, constraints, hyperparameters, justification

Cryptographic Hash Chain

Every layer is linked with SHA-256 hashes. Tamper-evident audit trail.

Regulatory Mapping References

ISAF maps your logged data to regulatory requirements. These are references to support compliance documentation — not certification.

EU AI Act

  • Article 10 - Data Governance
  • Article 11 - Technical Documentation

NIST AI RMF

  • MEASURE-2.2 - Evaluation metrics
  • GOVERN-1.1 - AI policies

ISO 42001

  • Section 8.4 - Control of externally provided AI

Colorado AI Act

  • SB 26-189 - Impact Assessment Documentation

CLI Tools

Inspect and verify lineage files from the command line

$ isaf inspect report.json

Inspect

View formatted report of lineage file with audit ID, timestamp, and logged layers.

$ isaf verify report.json

Verify

Verify cryptographic integrity of hash chain. Confirms data hasn't been tampered with.

$ isaf export-from-db lineage.db

Export

Export lineage from SQLite database to ISAF-compliant JSON format.

Research Foundation

ISAF is based on a technical methodology published as a preprint on Zenodo with a registered DOI

Published Whitepaper

The Instruction Stack Audit Framework (ISAF): A Technical Methodology for Tracing AI Accountability Across Nine Abstraction Layers

KC, S. (2025). Version 1.0. Zenodo. DOI: 10.5281/zenodo.18080355

This framework addresses the fundamental traceability gap in AI governance by providing technical specifications for documenting the full instruction stack from hardware substrate to emergent behavior. Includes a 127-checkpoint audit protocol and cryptographic verification methodology.

Read Full Paper →View All Research →
License: CC BY 4.0 | Released for citation, public scrutiny, and regulatory consideration

What ISAF Logger Is Not

ISAF is an evidence generation utility. Being explicit about what it does not do:

  • ✗Not a security scanner. ISAF logs lineage metadata. It does not scan code, detect vulnerabilities, or verify security controls. For security checks, see Developer Security.
  • ✗Not a compliance certification. ISAF generates evidence that can support compliance documentation. It does not certify compliance with any framework. Regulatory mappings are references, not certifications.
  • ✗Not a model behavior validator. ISAF does not validate model behavior, runtime safety, or regulatory compliance status. It logs what happened during training — not whether the training was correct or safe.
  • ✗Not a replacement for human review. ISAF evidence supports audit and review workflows. It does not replace human judgment, formal audit, or regulatory assessment.

Ready to Generate Lineage Evidence?

ISAF Logger is open source and free to use. Get started in minutes.

GitHub RepositoryEnterprise Support

Built by HAIEC — AI Security Validation & Audit-Evidence Platform