provenance

SchemaRouter introduces a lightweight schema-graph routing layer for heterogeneous agentic RAG systems, minimizing retrieval overhead by selecting only necessary tools and fields while maintaining answer accuracy and provenance attribution.

This paper introduces a claim-aware observability profile for autonomous scientific agents, emphasizing artifact lineage and audit relations to improve transparency and error detection in scientific workflows.

This paper presents centered residual signatures for data-free lineage verification of language model checkpoints, demonstrating high accuracy in distinguishing fine-tuned descendants from unrelated models with computational efficiency.

The article critiques the concept of AI-generated text and the effectiveness of watermarking for proving AI involvement, arguing that it may not capture human contribution and highlights ambiguity in defining AI slop.

This paper introduces Matrix, a deterministic causal-state layer for agentic workflows that records authority and fact dependencies, verifies completion evidence, and selectively invalidates stale work to ensure governed execution and auditability.

A blog post and accompanying tweet explore whether pre-generative-AI data becomes more valuable as the internet fills with synthetic content, discussing provenance, model collapse, and Anthropic's book scanning.

Introduces MAP-Graph, a provenance-aware shared memory layer for multi-agent workflows that filters by permissions, reranks by path trust, and gates risky actions, achieving 94.96% task success in controlled benchmarks.

Introduces the open-source project Semantica, which provides traceable decision-making basis and audit-support capabilities for AI Agents by building Context Graph / Knowledge Graph, suitable for scenarios with high explainability requirements such as finance and healthcare.

Anthropic announced it will watermark all Claude text output using an imperceptible watermark embedded in word choice rather than bytes; this analysis explains how that works and the layers where text watermarks can hide.

OMEM is a memory layer for AI agents that tracks beliefs over time, surfaces contradictions, and keeps provenance, aiming to replace static vector-store memory. The author is seeking early testers for this not-yet-production-ready tool.

Anthropic is adding invisible watermarks to all Claude text outputs and signed C2PA provenance metadata to supported files, in line with the EU AI Act transparency code of practice, with detection mechanisms to be detailed later.

This paper presents GPTKB 2.0, a large-scale disambiguated knowledge base derived from LLMs, containing 38.4M triples over 1.6M canonical entities. It offers a web interface for browsing, SPARQL/ natural language querying, and auditing fact provenance and disambiguation decisions.

A paper explores 'memory provenance laundering' in LLM agents, where long-term memory can turn untrusted observations into seemingly trusted context, and proposes preserving provenance through memory consolidation.

Semantica is an open-source tool that turns any data into a Context Graph, logging every AI agent decision with its full causal chain and W3C PROV-O provenance, enabling deterministic explainability for finance, healthcare, and government.

Us vs. Them is a CLI/library that analyzes git history to provide line-level provenance, showing which lines of text were written by humans versus AI agents under agentic editing.

Semantica is an open-source, self-hosted platform that converts data into a knowledge graph, records AI agent decisions, and provides full provenance and deterministic reasoning for explainable, traceable AI systems.

Presents SkillTrace, a multi-trace provenance auditing framework for LLM-agent skill reuse that extracts expression, implementation, and operational traces, achieving strong accuracy on a benchmark and enabling large-scale wild audits.

The article presents an agent architecture principle where arguments are looked up along a provenance chain (user_answer, instruction, pre_set_data, measured_data, prior_state) rather than generated, treating unknown fields as a valid state that triggers asking the user.

A survey of 'adversarial attacks for good', examining proactive protections applied across the visual content lifecycle to disrupt unauthorized AI automation and support accountability, covering five research communities such as privacy filters, unlearnable examples, generative safeguards, adversarial CAPTCHAs, and provenance mechanisms.

A developer introduces Proofpress, a prototype that embeds a portable revision/decision trail into Markdown and static HTML artifacts so agent-to-agent handoffs preserve context, with sidecar integrity evidence for DOCX.