# How Cloudflare enforces engineering standards using AI > [!question] Lingering Questions > - How often are AI findings wrong or disputed, and what does the appeal path actually look like in practice? The article gives volume metrics (250K deviations, 16K blocks) but no precision or false-positive data. > - Who can override an enforced MUST finding, and under what documented conditions? > - How does Cloudflare avoid its own standards corpus becoming bureaucracy for its own sake as it grows? > [!tip] Article Ideas > - Compare Cloudflare's "Codex" (owned RFC domains → structured extraction → lifecycle-gated agent enforcement) against Courtney's own AGENTS.md / skill-rule pattern — same shape at a much smaller scale. #articleseed > [!link] Cross Reference > - [[3. Resources/Permanent Notes/Review Scrutiny Should Scale With Consequence Not Production Method]] — same underlying instinct (review effort should track risk, not source), applied at the organizational-policy layer here versus the individual-review layer there. > - [[3. Resources/MOCs/AI Agent Safety MOC]] > [!quote] Vocabulary > - **RFC 2119 language** — the MUST/SHOULD/MAY vocabulary standard (IETF) that Cloudflare uses to give both humans and agents a shared severity grammar for standards statements. > - **Lifecycle states (draft/approved/enforced/deprecated)** — Cloudflare's model for how a standard moves from advisory to blocking, named here as the mechanism that keeps automated enforcement from becoming surprise bureaucracy. ## Bib capsule Cloudflare, "How Cloudflare enforces engineering standards using AI," The Cloudflare Blog, published circa early August 2026. ## Author's thesis AI code/spec/incident review only works as governance, not as a chatbot trick, when the organization writes and owns its standards first — organized into domains, owned by leads, written in RFC 2119 language, moved through lifecycle states from advisory to enforced — and lets agents consume structured, extracted statements rather than whole policy documents. ## Key arguments (≤3) 1. **Standards must be explicit and owned before an agent can enforce them.** Cloudflare's "Codex" organizes guidance into domains with named domain leads, written as RFCs — the article frames this as the actual hard work, with agent enforcement as the comparatively easy last step. 2. **Structured extraction beats whole-corpus prompting.** Rather than feeding an LLM the entire standards corpus, Cloudflare extracts SHOULD/MUST statements into JSON with stable slugs and metadata, letting agents retrieve only relevant rules — this keeps findings traceable to a specific, citable statement rather than a vague "this seems off." 3. **A lifecycle model (approved → enforced) is what makes blocking automation survivable.** Approved standards produce non-blocking findings; only enforced MUST violations can block a merge — this graduated rollout is presented as the mechanism that prevents automated review from reading as arbitrary or sudden to engineers. ## Claim atoms ### AI enforcement starts with governed standards - Claim: A review agent's output is only as trustworthy as the standards and ownership behind it — writing the standard has to precede writing the agent. - Evidence: Cloudflare created owned RFC domains with named leads before rolling out broad agent enforcement across code review, spec review, and incident review. - Candidate note: [[Automated Review Is Only As Accountable As The Standards Behind It]] ### Rule lifecycle protects adoption - Claim: Graduated enforcement (advisory before blocking) reduces the shock and backlash of introducing automated review. - Evidence: Approved RFCs generate non-blocking findings; only enforced RFCs with a MUST violation can block a merge. - Candidate note: [[Blocking Automation Needs A Lifecycle Not A Launch Date]] ### Structured extraction beats whole-corpus prompting - Claim: Extracting individual SHOULD/MUST statements into stable, slugged, retrievable units produces more traceable agent findings than handing an LLM the full policy corpus. - Evidence: Cloudflare's JSON extraction pipeline with stable slugs and metadata, explicitly contrasted against "dumping a whole standards corpus into an LLM." - Not separately promoted — treated as a supporting mechanism for the first claim above rather than a standalone durable claim. ## Permanent candidates - **Automated review is only as accountable as the standards behind it** — strong, atomic, generalizes past Cloudflare and past AI review specifically to any automated-enforcement system. Promote. - **Blocking automation needs a lifecycle, not a launch date** — strong, atomic, distinct claim about rollout mechanics rather than substance. Promote. ## Processing checklist - [x] Checked for existing Literature Note — none found - [x] Checked existing Permanent Notes for overlap on AI review/standards governance — found related-but-distinct review notes; linked rather than duplicated - [x] Connected into [[3. Resources/MOCs/AI Agent Safety MOC]] ## Links - [[3. Resources/MOCs/AI Agent Safety MOC]] — a concrete, at-scale example of governed-standards-plus-agent-enforcement architecture