jhcontext is a research programme and reference implementation for PAC-AI — a provenance-aware context protocol for auditable multi-agent AI systems, designed for EU AI Act compliance and adjacent regulatory regimes (FDA SaMD, MDR / MHRA, HIPAA / GDPR, NYC LL144).

PAC-AI specifies how AI agents exchange, audit, and prove the integrity of context — from specification, through SDK, into production deployment.

PAC-AI is being instantiated across regulated domains, each one a peer-reviewed paper in its own right.

Accepted (2026)

• PAC-AI: Provenance-Aware Context for Auditable Multi-Agent AI Systems — IADIS International Conference on Applied Computing & Intelligent Systems 2026 (camera-ready).
• From Explainability to Accountability: PAC-AI Envelopes for Auditable Oversight of Multi-Agent Clinical AI under the EU AI Act — AIiH 2026 Special Session, Imperial College London (LNCS proceedings).
• Auditable AI Assessment: Provenance-Aware Evaluation and Feedback for Student Work — AIET 2026 Special Session "AI-Based Evaluation and Feedback for Student Work", Zagreb.

Under review or in late preparation (described in general terms — most are under double-blind review):

• A cross-jurisdictional functional taxonomy for clinical-AI compliance stacks (US / EU / UK).
• A design-science instantiation of PAC-AI for AI-mediated hiring under the EU AI Act and NYC Local Law 144.
• An applied-economics cost model for EU AI Act non-compliance, and the mitigation value of provenance-aware context.
• A formal protocol paper (context fragmentation, semantic provenance, protocol semantics).
• A PRISMA-informed systematic survey of provenance, context, and accountability in multi-agent AI.
• A societal-frame paper on accountability by documentation in AI decisions (the reading version is published on jhcontext.com/research/pacai-accountability).

Full programme overview: jhcontext.com/research.

                    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

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}, optional situation + explanation boxes, and an administration.group classifier (Observation / Interpretation / Situation / Application). Directly SPARQL-queryable against the jh: 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:

pip install jhcontext

from 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()
)

The protocol is domain-general. The research programme has prioritised four regulated domains where the gap between deployed AI and contestable AI is most consequential:

• Healthcare — multi-agent clinical AI under the EU AI Act high-risk regime: temporal oversight, event logging, post-market obligations, HL7 FHIR-mapped evidence chains.
• Hiring & employment — AI-mediated hiring under EU AI Act Annex III §4(a) and NYC Local Law 144: negative-proof on protected attributes, sourcing neutrality, four-fifths disparate-impact audits.
• Education — AI-supported assessment with rubric-grounded grading: every feedback sentence binds to a rubric criterion and cites an evidence span.
• Finance & lending — quantified EU AI Act non-compliance exposure for European lenders, and the mitigation value of provenance-aware context.

jhcontext.com ・ jhcontext.com/research ・ YouTube ・ Substack ・ X ・ Threads

jhcontext is a research programme — reference implementation of PAC-AI, a provenance-aware context protocol for auditable multi-agent AI systems.