Skip to content

CRPv5 Improvement Roadmap

CRPv5 Is Shipped.
Here Is What Is Next.

CRP v5.1.0 is live: agentic positioning, the core quality layer (CDR/CDGR/CSO), the Safety Control Plane, the progressive SDK, pluggable storage backends, Gateway runtime, and Comply automation are all shipping today. This page tracks the improvements that make CRP the default trust infrastructure for AI agents: AIUC-1 certification alignment, offensive-security hardening, storage flexibility, and developer-experience polish.

Control-Evidence Mapping AIUC-1 Proof Point Browse Specs

1. Close AIUC-1 certification gaps

CRP already produces 80%+ of the control evidence AIUC-1 auditors verify. The remaining gaps are not technical limitations - they are accreditation, partnership, and scope decisions.

Gap Status Target
Accredited third-party adversarial testing Internal red-teaming is continuous; accredited attestation needed v5.2 - partner with AIUC-1 accredited lab
Third-party harmful-output validation Internal DPE testing runs on every call v5.2 - independent validation integration
AI cyber-misuse prevention No dedicated module v5.2 - add cyber-misuse detection stage to DPE
Catastrophic-misuse prevention No dedicated classifier v5.3 - evaluate with explicit weight-boundary controls
Insurance integration Not started v5.3 - explore AIUC-insured bundles

Certification is the floor

AIUC-1 certification is a point-in-time validation. CRP's architecture makes adversarial

defense, audit, and compliance evidence continuous by design.

2. Harden the Safety Control Plane

The Safety Control Plane is the central registry for all active safety rules. v5.x improvements make it more observable, tunable, and enforceable.

  • Rule versioning & rollback - every rule change is HMAC-signed and reversible.
  • Per-tenant rule isolation - SaaS deployments can run different rule sets per organisation.
  • Live rule simulation - test new rules against historical audit events before enabling them.
  • Safety-coverage heat map - visualise which risks are covered and which are explicitly out of scope.
  • Automated rule recommendations - suggest new rules based on audit patterns and benchmark failures.

See SPEC-033 and SPEC-034.

3. Expand pluggable storage backends

CRPv5.1 ships InMemory, SQLite, Redis, and S3 backends. v5.x improvements add:

  • PostgreSQL backend - for teams that want relational audit and CKF storage.
  • Azure Blob / GCS backends - cloud-native object storage beyond S3.
  • Encrypted local backend - AES-256-GCM encrypted files for air-gapped deployments.
  • Tiered backend router - automatically route hot, warm, and cold data to the right store.
  • Backend observability - latency, hit-rate, and cost dashboards per storage tier.

See SPEC-038 and SPEC-035.

4. Improve the progressive SDK

The SDK currently supports Level 0 (drop-in governance), Level 1 (ingest/ask), and Level 2 (depth/tools/fine-grained safety). v5.x improvements add:

  • Level 3 - Infrastructure control - manage Gateway, Comply, and Scan from the SDK.
  • Streaming safety events - subscribe to CRP-Safety-* headers in real time.
  • Multi-provider failover - automatic fallback across OpenAI, Anthropic, and local models.
  • Offline-first mode - queue calls and audit events when connectivity is intermittent.
  • Language clients - JavaScript/TypeScript and Go SDKs for non-Python stacks.

See SPEC-032.

5. Strengthen the Gateway runtime

CRP Gateway is the runtime product that turns every LLM call into a governed, audited event. v5.x improvements in flight:

  • 22-step OpenAI-compatible lifecycle - full request/response transformation parity.
  • Provider router with key vault - route by cost, latency, capability, and safety profile.
  • Visual console GA - real-time session tracing, risk dashboards, and audit exploration.
  • Edge deployment - lightweight Gateway worker for CDN-integrated inference.
  • Webhook destinations - stream audit events to SIEM, Slack, or custom endpoints.

See SPEC-016 and SPEC-043.

6. Advance the Semantic Task Layer (STL)

STL turns user requests into operation frames: RETRIEVE, COMPARE, ANALYSE, SYNTHESISE, GENERATE, VERIFY, CLARIFY, REVISE. v5.x improvements:

  • Add operation-specific safety budgets - VERIFY can be stricter than GENERATE.
  • Improve depth negotiation - auto-select D1–D5 depth based on task risk and context size.
  • Multi-operation plans - decompose complex tasks into bounded STL operations.
  • Goal-compass anchoring - prevent drift by re-checking task intent at each continuation.
  • Token-efficiency telemetry - prove <30% token cost vs. injection-equivalent prompts.

See SPEC-031.

7. Enhance CRP Scan remediation

CRP Scan finds ungoverned AI calls in codebases. v5.x improvements make it actionable:

  • Auto-remediation PRs - open pull requests with Gateway routing and policy stubs.
  • SARIF-native uploads - integrate with GitHub Advanced Security and other SARIF consumers.
  • Language expansion - TypeScript, Java, Go, Rust, and C# scanners.
  • IDE extensions - VS Code and JetBrains plugins for live governance hints.
  • Policy-as-code rules - let teams define custom governance rules in YAML.

See SPEC-013, SPEC-036, and SPEC-039.

8. Benchmark-driven quality gating

The Semantic Quality Benchmark (SQB) is the single gate that separates specification from proven system. v5.x improvements:

  • Publish SQB results - factual F1 and LLM-as-judge usefulness scores vs. v3 baseline.
  • Add adversarial SQB splits - measure robustness under prompt-injection and contradiction attacks.
  • Per-model scorecards - show which models perform best with CRP governance.
  • Continuous regression detection - run SQB on every release candidate.

See SPEC-026.

9. Standards-track acceleration

CRP is submitted to IETF, IANA, IEEE SA, and ISO/IEC JTC 1/SC 42. v5.x improvements:

  • Publish internet-draft updates - incorporate header-field registry feedback.
  • IANA registration - formalise CRP-* header fields.
  • Conformance suite v2 - automate Level 1/2/3 conformance testing.
  • Public test vectors - let anyone verify CRP compatibility independently.

See SPEC-014 and Standards Track.

Release timeline

Release Focus Expected
v5.1.0 Core quality layer, safety surface, SDK, storage backends Current
v5.2.0 AIUC-1 accredited testing integration, cyber-misuse DPE stage, PostgreSQL backend Q3 2026
v5.3.0 Catastrophic-misuse classifiers, insurance partnerships, language SDKs Q4 2026
v5.4.0 Gateway visual console GA, edge deployment, IDE extensions Q1 2027

Contribute

CRP is open source under the Elastic License 2.0. If you want to help close a gap, start with the conformance suite or open an issue on GitHub.

Want to influence the roadmap?
Join the standards discussion or request an enterprise feature.
Open an Issue Contact Standards