model-provenance

Toolkit to assess and determine model provenance

Local evidence, diagnosis, and control loop for prompts, checkpoints, and AI agents.

[USENIX Security 2026] Attesting Model Lineage by Consisted Knowledge Evolution with Fine-Tuning Trajectory

AetherGuard AI is presented as a holistic AI Trust & Integrity Gateway that expands the typical AI gateway paradigm to offer real-time semantic inspection, cryptographic accountability, responsible AI compliance, and robust operational governance. The system operates as a transparent reverse-proxy between LLM clients and providers.

Offline-first evidence and verification framework for AI model artifacts, transformations, runtime identity, and provenance.

Runtime guardrails, behavioral evaluation, and signed fleet admission for Hermes Agent

Open knowledge base for Provenance Engineering: data lineage, identity governance, model provenance and software inventory.

Decentralized, poisoning-resistant distribution for ML models and datasets. Publishers sign once, untrusted mirrors distribute, and any consumer verifies the exact signed bytes with full provenance.

Validate model and release claims against small provenance envelopes with redacted output.

Anchor MLflow artifacts to Bitcoin. One line to production.

An offline-first computer vision integrity and security assurance platform for auditing datasets, model artifacts, and inference outputs across ML pipelines.

Static analysis toolchain for neural network weights - inspect, diff, fingerprint, and scan model checkpoints without running them. Ghidra for models.

A protocol for collecting, fusing, and assessing multi-source evidence of model derivation without treating audit signals as automatic legal or royalty verdicts.

Anchor W&B artifacts to Bitcoin. Zero-touch provenance for ML experiments.

AIBOM toolkit: CycloneDX 1.5 validator, JSON schema, hash utilities for AI BoM generation, validation, and supply-chain transparency.

Code and data for "Does a Violation Record Run in the Family? Inherited Rule-Breaking and Lineage Linkage in Open-Weight Language Models" (IEEE Access submission). Archived at doi:10.5281/zenodo.23378391

Signed, offline-verifiable records for AI models. Open standard, Apache 2.0.

Verifiable fine-tuning on 0G: each model gets a signed, on-chain-anchored birth certificate anyone can verify.

Which AI models does your code use, whose weights are they, and which model wrote your commits? Offline CLI with a --deny CI gate and CycloneDX 1.6

LLM distillation detection: keyed fake facts survive paraphrase laundering (10.4σ) where token watermarks fail (4.4σ)