The loop, demonstrated — a free walkthrough
One issue, end to end.
The doctrine on this estate compresses to one sentence: AI drafts, humans approve, machines and customers consume, and the system reports what is missing. Here is that sentence made concrete — a single support ticket riding the whole loop, with about seven minutes of human judgment in its entire life.
The trace comes from the toolkit's worked example universe (Meridian Systems, a fictional company built to be worked through); the system shape is the real product.
Act 1 — a resolved case becomes governed knowledge.
A customer's VPN "connects for a second and then kicks me out" after a macOS update. The agent fixes it and closes the case with a one-line wrap-up note. That sentence is the entire capture tax — about twenty seconds. A synthesis prompt runs against the thread and emits one complete, typed candidate — and refuses to guess what the thread never stated:
A human curator opens the gate: dedupe, grounding, applicability. They confirm the flagged version against the case logs and add the one thing the draft could not know — the fix does not apply to personally-owned, non-MDM devices. Light-touch approval, edit-distance 1, about four minutes. The candidate is promoted, validated, and only now may any delivery surface serve it. One append-only row records the decision:
Act 2 — the world changes, the article keeps up.
Three months on, a macOS point release moves the settings path the fix depends on. A product-change signal fires and the update prompt produces a surgical DIFF — not a rewrite:
The gate applies the meaning-change test: is truth changing, or only its address? Users on the older OS still need the old path, so the DIFF adds the version-scoped step instead of overwriting it. Accepted — about two minutes. Without that gate, the other timeline runs: a confident answer engine citing stale instructions at every customer the day the OS ships.
Act 3 — what one governed article is worth.
With illustrative numbers: this symptom was hitting the assisted queue ~70 times a month at ~$18 a loaded answer — roughly $1,260/month to re-answer a known question. Validated, the article serves the same issue through the help center, the agent-assist panel, and the answer bot at a unit cost measured in cents; within two months the assisted residue is single digits, mostly the edge case the curator fenced off in Act 1. Assisted time-to-resolve rises — by design; the easy work left.
And the sixth channel is already implied: a customer's own AI agent consuming the same validated, cited, applicability-filtered article — where the metadata the curator verified stops being hygiene and becomes the product.
The point.
Seven minutes of human judgment, applied only where judgment is the actual job — two gate decisions, one boundary the machine could not know, two log rows. Everything else was built once and reused: the typed templates the machine drafts into, the grounded prompts, the checklists, the logs, the workbook the numbers roll into.
One caution, because the walkthrough makes it look easy
It worked because it was sequenced — standards ratified first, the gate stood up second, the prompts turned on last. Run the acts in that order. Running them in reverse is how you ship confident, cited, wrong answers at scale.
The full walkthrough — and every working file it references: the templates, the prompts, the gate checklists, the governance logs, the 16-tab measurement workbook — ships in the Contextkeeping Toolkit.