Lessons 89.1–89.4 built the machinery of scale: quality gates, cost guardrails, runbooks, red lines. Now the habit that keeps it alive — because every system drifts. Prompts bloat, vendors change prices, exceptions creep up, owners rotate. Without a cadence, drift wins quietly. With one thirty-minute meeting a week, you catch it early and fix one thing at a time.

The vocabulary, in plain language

• Operating cadence: the fixed rhythm of review — weekly for working flows, monthly for costs and strategy. Cadence beats intensity; a short meeting every week outperforms a heroic offsite twice a year.
• Decision log: the written record of what was decided, by whom, and why — so next month nobody re-argues it from memory.
• Experiment: one bounded change with a prediction and a check date — "tighten the ambiguity threshold; expect exceptions down 3 points by the 24th." Not a vague "improve triage."
• Owner: the named person accountable for the workflow and for the experiment. Shared ownership is shared neglect.
• Next constraint: the single bottleneck that most limits the operation right now (Lesson 84.2). You fix that one before optimizing anything else.

The mental model: a doctor's weekly rounds. Same charts every visit (outcome, quality, cost, exceptions), one patient complaint heard seriously (customer signal), one prescription at a time (experiment), everything charted (decision log). No rounds, no recovery.

The weekly agenda (30 minutes, no more)

Seven signals, one decision. Timebox it or it becomes a status ramble:

| Min | Signal | Question |

|---|---|---|

| 0–4 | Outcome | Did the workflow deliver its job — volume, completion time, customer result? (OPS-SCORECARD, 84.5) |

| 4–8 | Bottleneck | Where did work wait or pile up — queue age, handoffs? What's the next constraint? |

| 8–12 | Quality | Rejections, rework, audit findings? (QUALITY-STANDARD, 89.1) |

| 12–16 | Customer signal | One ticket, reply, or metric customers actually felt? (Support → product loop, 87.5) |

| 16–20 | Cost surprise | Unit cost drift, alerts, vendor changes? (GUARDRAILS, 89.2) |

| 20–24 | Exceptions | Queue depth, aging items, top failure case? (contract 88.4) |

| 24–30 | One improvement | Single experiment: what, owner, prediction, check date. Log it. |

Sources are the artifacts you've already built: scorecard, quality log, cost guardrails, review queue, automation reviews. The meeting reads instruments; it doesn't invent opinions.

Exercise: start the review

Create OPERATING-REVIEW.md:

# OPERATING-REVIEW.md — [Operation] — week of ___

## Signals (numbers + one line each)
- Outcome: ___ | Bottleneck / next constraint: ___
- Quality issue: ___ | Customer signal: ___
- Cost surprise: ___ | Exceptions: ___ (queue depth ___, oldest ___)

## Decision log
- ___ decided by ___ because ___ (date ___)

## Experiment (exactly one)
- Change: ___ / Owner: ___ / Predict: ___ / Check: ___

## Standing agenda (30 min weekly)
- Outcome 4′ → bottleneck 4′ → quality 4′ → customer 4′ → cost 4′ → exceptions 4′ → one improvement 6′
- Roles: facilitator ___ / notes+decision log ___ / workflow owner ___

Worked mini-example — ByeBuy classroom ops week: outcome 2 lessons drafted on time; bottleneck — artwork review waited 3 days (next constraint: reviewer capacity); quality — 1 rejection (unsourced claim, corrected + checklist note); customer signal — two readers asked the same setup question (→ FAQ candidate); cost — model spend +18% after a prompt change (experiment: trim context, predict back under cap by next review); exceptions — 4 triage items, oldest 2 days, all routed. One experiment only: the context trim, owned, dated. Next week reads the result.

Finish line: a dated OPERATING-REVIEW.md with all seven signals, one logged decision, and exactly one experiment with owner and check date — plus the standing agenda.

Verify quickly: count the experiments. More than one means nothing will finish. Zero means the meeting was theater. One, with a name and a date, means the system improves.

Common failure mode: the drift meeting — forty-five minutes of anecdotes, no numbers, three "let's all try harder" action items with no owners. It feels productive and changes nothing. The agenda above exists to make that meeting impossible.

Check your understanding

1. Why exactly one experiment per week — what breaks with three? 2. What does the decision log prevent that memory can't? 3. Which artifact feeds each of the seven signals?

Next — Part XVI completion

You now hold the full scale toolkit: operation maps and contracts (84), content and marketing ops (85–86), support memory (87), careful automation (88), quality, cost, runbooks, red lines, and review (89). Assemble the OPERATIONS-PLAYBOOK folder — operation map, automation contract, scorecard, review-queue design, permissions matrix, runbook, cost guardrails, operating-review agenda for one real workflow — and be able to explain what runs, why, who owns it, and when it stops. Part XVII turns it into your scoped operating plan.