> Markdown copy of [https://aimaturitymodels.com/models/sdlc/deep-dive/d6/](https://aimaturitymodels.com/models/sdlc/deep-dive/d6/), generated at build time: a rendering, not a source of truth. Map: [/models/llms.txt](/models/llms.txt). Site map: [/llms.txt](/llms.txt).

# AI-Native SDLC Maturity Model — D6

## Architecture governance

Every codebase has architectural taste, even if it only lives in one senior engineer's head. This dimension asks whether that taste — the patterns to follow, the ones to avoid, the reasons behind both — is something the organization actually has, in a form anyone (or any agent) can find and act on, or something that only exists as tribal memory.

### Where most organizations start (Nascent)

Architectural standards exist only as tribal knowledge or informal convention, enforced inconsistently through manual review — meaning quality depends entirely on which reviewer happens to look at a given change. The first real step is externalizing that judgment: documented decisions, constraints, preferred and prohibited patterns, with a named owner for keeping the record current.

### Where the real gains happen (Modeled → Integral)

The meaningful shift starts with encoding the most common and consequential violations as deterministic checks — linters, structural tests, CI gates — so consistency stops depending on which reviewer is on duty. From there, the bigger change is converting those checks from after-the-fact gates into something agents can read and apply *during* generation, with remediation guidance attached, so violations get caught before code exists rather than after.

> **Worth knowing about this step:** the shift from encoded-but-partial checks to comprehensively agent-readable, generation-time architecture is a genuinely larger lift than its neighboring steps — flagged as a candidate "lumpy" transition, reviewed and confirmed rather than smoothed over. David's own review: "Steep but inescapable given current understanding, and maps to a correspondingly large reward for the cost."

### What the top of the curve actually looks like (Telemetric)

At full maturity, agents don't just follow the encoded standard — they participate in evolving it. When an agent hits a pattern the rules don't cover, or notices a rule that's become a recurring obstacle, it proposes a change, with evidence, for human review. The architecture and the enforced standard co-evolve, with agents as contributors to that evolution, not just subjects of it.

### Why this dimension matters

D6 sits over the seam between what gets built (D4, D5) and what gets verified (D7, D8) — it's the dimension that decides whether "good architecture" is a matter of luck-of-the-reviewer or something the system itself actually protects.

---

*Drafted from the SDLC model's real locked content — including the transition and verification notes now folded into `ai_native_sdlc_maturity_model.md` itself (2026-07-27). D6's C→D "lumpy transition" question, named in this model's own Open Items, was reviewed and resolved 2026-07-24 — no open review flag remains.*