The Capability Maturity Model Integration (CMMI) is the canonical staged-maturity framework for software engineering. Developed by the Software Engineering Institute at Carnegie Mellon and now maintained by ISACA, CMMI has been applied to thousands of engineering organizations since the mid-1990s.
CARL borrows CMMI's structural form — named levels, gated progression, staged representation — while targeting a different question and a different evaluation method. CARL is not a 1:1 CMMI staged map: its ladder runs 0–6, adding a manual floor (L0) below CMMI's Level 1 and two aspirational business-operating rungs (L5, L6) above the top of CMMI.
Level mapping
CARL's middle rungs echo CMMI's intent, but the endpoints diverge — L0 and L5/L6 have no CMMI peer:
| CARL | CMMI Staged |
|---|---|
| L0 Manual | (no peer — below Level 1) |
| L1 Scripted | Level 1 Initial |
| L2 Tested | Level 2 Managed |
| L3 Reviewed | Level 3 Defined |
| L4 Agent-driven | Level 4 Quantitatively Managed |
| L5 Creates a business | (no peer — beyond CMMI) |
| L6 Runs a business | (no peer — beyond CMMI) |
The middle mapping is approximate, not exact. A team already speaking CMMI vocabulary can read a CARL report's L1–L4 rungs without much translation, but CARL's floor and ceiling are deliberately outside the CMMI frame.
What CARL adopts from CMMI
- Staged representation. Levels are ordered; a subject must pass all prior levels to claim a higher one. Shared structural idea.
- Gated progression. A level is either achieved or not — no partial level scores.
- Named levels with normative intent. L0–L6 are not arbitrary ordinals; each has a stable interpretation, from a manual repo (L0) up through an agent-driven codebase (L4) to the aspirational create/run-a-business rungs (L5, L6).
What CARL does differently
Codebase-native, not process-native
CMMI asks about the process: what steps do you follow, what artefacts do you produce, what reviews do you conduct? CARL asks about the codebase: what files are present, what configuration is enforced, what signals are instrumented?
The difference matters because agents don't execute processes. They read files, run commands, observe outputs. CARL measures exactly those things.
Machine-checkable by default
CMMI appraisals require trained CMMI Institute appraisers conducting interviews, reviewing artefacts, and producing appraiser-subjective findings. The cost is substantial — often six-figure engagements — and the variance across appraisers is real.
CARL's default assessment is a deterministic tool reading a repository. 80%+ of criteria are deterministic, reproducible, and reportable without an appraiser. AI-assisted criteria carry disclosed variance.
Codebase-scope (repository / application / organization)
CMMI assesses an organization's engineering process. CARL can assess a single repository, a single application within a monorepo, or an organization aggregating repositories — and produces comparable, scoped results at each level.
Eight pillars specific to agent-era work
CMMI's process areas (Requirements Management, Project Planning, Configuration Management, etc.) reflect enterprise engineering concerns of the 1990s and early 2000s. CARL's eight pillars — Documentation, Code Quality, Testing, Build, Observability, Security, Task Intake, Delivery & Measurement — reflect what modern agent-assisted development actually depends on.
Vendor-neutral standard, no appraiser gate
CMMI is maintained by ISACA, which sells appraisal training, certifies
appraisers, and licenses the model. CARL is published and maintained by
Wentzel.ai; the standard and reference implementations are proprietary today
(see License), but no appraiser certification or
paid training is required to self-assess, claim a level, or read the spec at
carl.wentzel.ai.
Is CARL replacing CMMI
No. A large enterprise running a formal CMMI appraisal every few years should keep doing so if it serves its purpose. CARL is better suited to:
- Measuring the specific technical substrate that agent-assisted development depends on.
- Running assessments frequently — every PR, every release, every 90 days — rather than as a periodic exercise.
- Producing machine-comparable results across organizations and over time.
The two can coexist. A CMMI Level 3 organization often clears CARL L3 Reviewed on its flagship repositories but falls short on newer ones; a CARL L3 repository inside a CMMI Level 2 organization is evidence that the organization has at least one codebase operating at a mature level.
Related
- The 7 levels — the full interpretation of each CARL level (the 0–6 ladder).
- Assessment methodology — how CARL assessors produce results.
- vs. Microsoft's model — the organizational-adoption counterpart to CARL's technical substrate.