Context Governance Benchmark (CGB)
June 10, 2026 · View on GitHub
Status: v1.0-draft — pre-review. Not yet reviewed by the external review board; do not cite as a released standard. See GOVERNANCE.md.
CIS Benchmarks define what a hardened server is. The Context Governance Benchmark defines what a governed context pipeline is: a versioned, tool-neutral catalog of measurable requirements for the layer that feeds data into AI models — retrieval, compression, memory, tool output, shared knowledge.
Why this exists
Every team wiring agents into real codebases and real data answers the same questions ad hoc: Can secrets reach the model? Who spent those tokens? What exactly entered the context window? Can we prove any of this to an auditor? This catalog turns those questions into named, measurable controls, so that security reviews, procurement and vendor self-assessments can stop reinventing the checklist.
What is in scope
A context pipeline is every component between data sources and the model-bound prompt: file and shell access layers, retrieval and search, compression and summarization, session and long-term memory, connectors, and the budget/audit machinery around them. The model itself, prompt-level jailbreak resistance and agent reasoning quality are out of scope — other work covers those.
Structure
| Domain | Controls | |
|---|---|---|
| 1 | Sensitivity & Redaction | CGB-1.1 … 1.6 |
| 2 | Provenance & Integrity | CGB-2.1 … 2.5 |
| 3 | Budget & Resource Control | CGB-3.1 … 3.5 |
| 4 | Audit & Evidence | CGB-4.1 … 4.6 |
| 5 | Access Scoping | CGB-5.1 … 5.5 |
| 6 | Lifecycle & Retention | CGB-6.1 … 6.5 |
32 controls, three levels (Basic / Hardened / Audited), four maturity grades (C1–C4). Scoring: SCORING.md.
Neutrality
Controls name capabilities, not products. No vendor's terminology — including that of the benchmark's initiators — appears in any control. This is enforced by CI (tools/lint_neutrality.py). Vendors document their conformance in separate self-assessments using assessment/TEMPLATE.md; honest gaps are a feature, not a bug.
Assessing a pipeline
- Copy assessment/TEMPLATE.md.
- Work through all 32 controls; apply each control's measurement method.
- Score per SCORING.md; publish with version, date and the self-assessment label.
Contributing
Issues and merge requests are welcome — control wording, measurement methods, missing domains. Substantive changes follow the RFC-light process in GOVERNANCE.md. Reviewer seats for v1.0-final are open to practitioners in security, compliance and platform engineering.
License
Text: CC BY 4.0 — cite, adapt, redistribute with attribution. Cite as: Context Governance Benchmark v1.0-draft, CGB-x.y.