Field notes · essay
Boring core, smart edge: the operating model behind the slogan
"Boring core, smart edge" fits on a sticker, which is exactly its weakness. Slogans travel; operating models don't. The AI knowledge programs I've watched stall rarely stalled on the technology. They picked a capable model, wired up retrieval, and demoed well. Then, a few months in, the answers started drifting, the review queue started aging, and nobody could say who had approved what. The model held up fine. The rhythm didn't.
Here is the operating model the slogan is short for. None of it is clever. That's the point.
The core is boring on purpose
At the center of a governed knowledge operation sits a deterministic, versioned system of record: every article typed, owned, tagged, lifecycle-dated, with its history intact. In the reference implementation I run, the record is a git repository, and every accept, validate, and archive is a commit made at the gate. Boring here means diffable, auditable, and reversible. It means "what did this article say on the day the customer read it?" has an answer.
Around the core, the edge is smart and probabilistic. AI drafts candidates from resolved cases on the way in, assembles cited answers on the way out, clusters the gaps, scores the candidates. The edge should be as capable as this quarter's models allow, and it will be replaced by next quarter's. The core shouldn't notice; the AI layer is a disposable projection of the record, never the record itself.
Between the two runs one rule: nothing probabilistic writes here without passing the gate.
Three governed loops cross that boundary. Capture synthesizes knowledge from work. Curation decides what enters the record and when it leaves. Delivery serves cited answers to people and machines, and logs every question it couldn't answer as a capture request. All three loops use the edge. Only curation writes to the core.
Four roles, budgeted in seconds and minutes
The governed loop runs on effort budgets, not job descriptions.
- Agent. Flags a resolved case worth keeping; AI drafts, the agent judges. Budget: thirty seconds per capture. Above that, the capture tax comes back, and an agent measured on throughput will do the math.
- Curator. Runs the gate: approve, edit-approve, merge, or reject with a reason code. Budget: three minutes per clean approve, which is roughly twenty an hour. Two or three trained curators are enough to pilot.
- Knowledge owner. Owns the record itself: lifecycle states, review cadences, the terminology standard. A standing role, not a project.
- Executive sponsor. Owns one number, blended cost per answer, plus the authority the gate borrows. One review a month.
The budgets are design constraints. Everything else in the model is arranged so they hold. Compare the usual alternative, the review committee. A committee has no effort budget: it meets on a calendar, the queue ages between meetings, and people learn to route around it. A curator with a three-minute budget and a one-business-day publish clock can't quietly become a bottleneck; the gate either keeps pace or visibly doesn't, and you fix it that week.
The sponsor role isn't ceremonial either. The gate holds because someone with budget authority says it holds. That's what ratifying a governance policy actually means.
Three gates, one log that never forgets
Three gates cover a record's whole life. The new-candidate gate asks whether something enters. The update-diff gate asks whether a change stands, and a human reviews the before-and-after, not just the after. The retirement gate asks whether something leaves honestly. Every pass ends in exactly one disposition (approve, edit-approve, merge, or a coded rejection), and every disposition lands in an append-only decision log.
Append-only is the most underrated choice in the model, and it's what makes AI drafting safe. A gate whose history can be edited can't be trusted, by an auditor or by you. With the log, every AI draft has a traceable fate: who decided, when, with which reason code, and how much they changed. Reason codes turn rejections into telemetry instead of friction. Edit distance, how much a curator changed a draft before approving it (a benchmark credited to Michael Iantosca), trends down as drafting improves.
The log also makes the edge swappable in practice, not just in principle. When a new model replaces the old one, as it will, the decision log tells you whether it drafts better or worse, on the same scale you measured its predecessor on. No vibes required.
Four clocks
Knowledge operations fail on rhythm before they fail on tooling, so the governed loop runs on four clocks:
- Capture at resolution: the case is flagged while it's still warm.
- Publish within one business day of capture.
- Review on volatility tiers: evergreen, medium, and high-volatility content each get their own re-verification cadence, and a verified date moves only when a human actually re-reads.
- Placement quarterly: the program is re-scored against the maturity model, so the roadmap answers to evidence.
Each clock guards against a specific decay. Miss the first and capture happens from memory. Miss the second and the review queue turns into a warehouse. Miss the third and the core rots while still looking authoritative, which is worse than looking stale. Miss the fourth and the roadmap drifts toward whatever was loudest last month.
Sequencing is the hard part
Programs rarely fail on intent; they fail on sequencing. The order is the least glamorous element of the operating model and the most decisive. Foundations first: the record, the gate, the roles named. Then a pilot, with templates live, prompts wired to real ticket exports, and two or three curators running the gate. Then scale on what the pilot proved. A 90-day arc, not a big bang.
The common failure runs that order backward: the smart edge first, because it demos, and the boring core later, because it doesn't. The demo is great. Then the answers drift, because there was never a governed record for them to stand on, and the program gets blamed on the model.
None of this is about the model. It's discipline: budgets, gates, clocks, and order. That discipline is unglamorous, which is why it gets skipped, and durable, which is why it wins. Models come and go at the edge.
The boring core compounds.
The roles, gates, clocks, and the one number the program answers to are laid out for the budget owner in the operating model. Why the record needs a gate at all is argued in the doctrine, and what blended cost per answer measures is Answer Economics.
KCS® is a service mark of the Consortium for Service Innovation™. This essay reflects an independent, KCS-informed methodology and is not affiliated with or endorsed by the Consortium. The edit-distance benchmark is credited to Michael Iantosca and draws on his published papers at thinkingdocumentation.com/downloads.