Every page here is also a document. Use your browser's print or save-as-PDF and you get the same PDF that goes out by email, which is why these are laid out like sheets of paper.
EZO CONTENT PIPELINE · JULY 2026

How the QA layer collapses the review gauntlet

Same standards. The machine enforces the rules; humans keep the judgment.

TODAY median 16 days · 11 steps
cycle time, tasks completed last 12 months: median 16d, mean 35d, p90 79d
1●Outline64
2●Research and SERP25
3●Draft73
4●Editing31
THE REVIEW GAUNTLET - 6 STAGES
5●Content review37
6●SEO review49
7●PMM review57
8●Rubab's review49
9●Rob's review47
10●VP review36
11●Publish35

6 humans re-checking rules a machine can check.
Counts = Asana tasks per stage.

PROPOSED hours · 5 steps
the SOPs already define the rules; the rule checker enforces them
1○Structured briefSOP template + topic + demand evidence
2○AI draftborn inside the SOP structure - right sections, right char counts
3●MACHINE QA GATE
INstructured draft page (JSON) · SOP ruleset · product-fact KB (137 facts, seeded from EZO's own published content, verified before the system trusts it)
DOfind → fix → verify → re-check  ·  the rule checker finds problems; fixer proposes minimal patches; applier enforces that only flagged fields change; an independent verifier accepts or rejects each patch (factual edits must cite a fact); the re-check confirms
OUTscorecard (before/after, 0-100) · patched page · exceptions-only human queue, each with a reason · journal of every change with rollback
live eval: catch 100% · fix 86% · unjustified edits 0% · verifier true-negatives 100%
4○One human reviewPMM/SME judgment on the ~20% a machine can't check
5○Publish + verifylive URL confirmed, not assumed
INSIDE THE CHECK - THREE PASSES, CHEAPEST FIRST
● The rule checker - deterministic lint
char ranges, section order, element counts, CTA rules, slug/meta, banned and compliance phrases. Any CRITICAL = auto-fail.
→
● The AI reviewer - two-model editorial (fix + verify built ● · 5-dimension scoring proposed ○)
fixer + independent verifier; every factual edit cites an approved fact. Scored 1-5 on accuracy, clarity, voice, intent, originality.
→
○ The human reviewer - one sign-off
judgment on the 20%: accuracy edge cases, positioning calls. Reviews the exceptions, not the checklist.
○ Proposed pass bar (mark it up): overall = 60% machine + 40% editorial · pass = overall ≥80 AND no CRITICAL AND product accuracy ≥4/5
Why it can be trusted: nothing outside a flagged spot changes · verifier ≠ author · every change has a receipt (before/after + fact citation + verdict) · one-command rollback · 75 tests green, live on the Anthropic API
● built, runs today   ○ proposed, mark it up