A provenance-aware context protocol for multi-agent AI systems, designed for EU AI Act compliance. jhcontext defines how AI agents exchange, audit, and prove the integrity of context — from specification to production deployment.
jhcontext-protocol
(JSON-LD specification)
│
▼
jhcontext-sdk
(Python SDK on PyPI)
╱ ╲
▼ ▼
jhcontext-usecases jhcontext-crewai
(in-memory POC) (AWS production)
~25 ms, no infra Lambda + DynamoDB + S3
| Repository | What it is | Start here |
|---|---|---|
| jhcontext-protocol | JSON-LD specification (v0.5). Defines the envelope structure, UserML-correct SituationalStatement shape (Heckmann five-tuple mainpart), forwarding policies, and W3C PROV provenance mappings. | jhcontext-core.jsonld |
| jhcontext-sdk | Python SDK. EnvelopeBuilder, ForwardingEnforcer, StepPersister, PROV graph builder, PII tokenization, audit functions, FastAPI server, and MCP server. | pip install jhcontext |
| jhcontext-usecases | Lightweight proof-of-concept. Healthcare (Art. 14 temporal oversight) and Education (Art. 13 negative proof) scenarios with a 7-benchmark suite. Runs in ~25 ms, no infrastructure needed. | python -m usecases.run |
| jhcontext-crewai | Production deployment on AWS. CrewAI multi-agent flows for Healthcare, Education, Recommendation, and Finance with Chalice Lambda API, DynamoDB, and S3 storage. | docs/architecture.md |
An envelope is a context container that travels between AI agents. It carries:
- Semantic payload — a SituationReport of atomic UserML SituationalStatements (Heckmann 2005), each with a five-tuple
mainpart {subject, auxiliary, predicate, range, object}, optionalsituation+explanationboxes, and anadministration.groupclassifier (Observation/Interpretation/Situation/Application). Directly SPARQL-queryable against thejh:vocabulary. - Artifacts tracking every computational product (model outputs, embeddings, tool results)
- Forwarding policy with monotonic enforcement — once set to
semantic_forward(HIGH-risk), raw context is permanently filtered - W3C PROV graph linking entities, activities, and agents across the pipeline
- Cryptographic proof via URDNA2015 canonicalization, SHA-256 hashing, and Ed25519 signatures
- Privacy and compliance blocks for PII tracking and regulatory metadata
Six auditable operations, each demonstrated end-to-end in the usecases and crewai repos. Every verifier is a thin wrapper over a SPARQL query against the recorded SituationReports:
| Pattern | EU AI Act | What it proves |
|---|---|---|
| Temporal oversight | Art. 14 | A human reviewed AI output after the recommendation, with verifiable timestamps |
| Negative proof | Art. 13 | Protected attributes (identity, disability) were absent from the decision chain |
| Workflow isolation | Art. 13 | Parallel workflows (e.g., grading vs. equity) shared zero PROV entities |
| Integrity verification | General | SHA-256 hash and Ed25519 signature over canonical JSON-LD remain valid |
| Rubric grounding | Art. 12 + Art. 86 | Every LLM feedback sentence binds to a rubric criterion and cites an evidence span in the student text |
| Multimodal binding | Art. 12 | Audio / image / video artifact citations resolve to the exact region in the referenced source |
pip install jhcontextfrom jhcontext import EnvelopeBuilder, RiskLevel, observation, interpretation
# Build a SituationReport — a flat list of atomic UserML SituationalStatements
payload = [
observation("user:alice", "temperature", 22.3,
range_="float-degrees-celsius",
source="sensor:thermostat-01"),
interpretation("user:alice", "thermalComfort", "comfortable",
range_="uncomfortable-neutral-comfortable",
confidence=0.92),
]
env = (
EnvelopeBuilder()
.set_producer("did:example:agent-1")
.set_scope("healthcare")
.set_risk_level(RiskLevel.HIGH) # auto-sets forwarding_policy=semantic_forward
.set_human_oversight(True)
.set_semantic_payload(payload)
.sign("did:example:agent-1")
.build()
)jhcontext.com ・ YouTube ・ Substack ・ X ・ Threads
jhcontext is a research project — reference implementation of PAC-AI, a provenance-aware context protocol.