Back to Agents That Account for Themselves
working paper2026Graph Kernel paper

Deterministic Provenance Engines for Autonomous Agent Systems

The Graph Kernel paper describes a deterministic provenance engine for language agents. Rather than treating retrieval as a bag of similar snippets, it produces signed, policy-governed context windows with evidence boundaries that downstream agents can audit.

Paper workspace

Live draft structure

working-draft

Artifacts

Draft PDF

Graph Kernel v2 draft PDF, public for reading as a current working artifact.

Open artifact

Editable source

Draft PDF exists. It should stay editable as Graph Kernel and live knowledge graph work converge.

Source anchors

Comp-Core/docs/GRAPH-KERNEL-PAPER-V2.md

Comp-Core/docs/latex/graph-kernel-paper-v2.tex

Comp-Core Graph Kernel Rust service

Method tags

provenancegraph retrievaladmissible contextdeterminism

Ingest intersections

graphretrievalprovenanceagentmemory

Status

Paper drafted; system running.

Key claims

01

Context should be admissible evidence, not unbounded memory paste.

02

Graph traversal can answer multi-hop structural questions better than flat retrieval.

03

Provenance tokens prevent agents from inventing authority.

Public reading note

Paper drafted; public page can safely describe the architecture.

Standard skeleton

What this paper must keep proving

Schema

problem

Agents treat retrieved context as authority even when the retrieval path is not inspectable.

method

Produce signed, policy-governed context windows from graph traversal instead of loose snippets.

implementation

Rust graph kernel, admissibility tokens, policy slicers, deterministic traversal.

data

Personal and project graph corpora with strict provenance boundaries.

evaluation

Multi-hop relevance, latency, determinism, evidence completeness, and comparison against flat retrieval.

references

GraphRAG, retrieval-augmented generation, knowledge graph traversal, provenance systems.

openQuestions

How much of the live knowledge graph paper should merge into this paper versus remain a companion.

Checkpoints and references

Proof chain

implementationproven

Graph Kernel implementation

Comp-Core Graph Kernel Rust service

The claimed retrieval/provenance substrate exists as a running implementation.

experimentpartial

Multi-hop comparison

Graph Kernel paper benchmark tables

Result exists in the paper draft and should remain tied to the exact benchmark snapshot.

external-referencepending

External GraphRAG comparison

related-work slot

Needs current citations as the field changes quickly.

Reference links

Live Knowledge Graphsextends

GraphRAG and knowledge graph retrievalneeds-citationrelated-work slot