The Operating Cadence
A strategy on one page and forty people do not run themselves. This part closes Phase 5 with the operating system of the company: the rhythm of meetings from daily stand-ups to the yearly plan, quarterly goals tied to the strategy, a weekly operating review built on input metrics, a decision table checked by a script for self-approval and missing owners, a reviewer roster that keeps the gate independent as the team grows, the board's role, and the incident that tested all of it. By the end, Gopala is a CEO with a company that can make good decisions when he is not in the room.
decision_rights.py and gate_roster.py were run (real output below). The decision table, people, updates, metrics, meeting hours and the incident are the story's scenario. Management methods are cited from their sources. Gopala and his company are a teaching story. Disclosure: Claude, made by Anthropic, helps write this site.
1. Mission log: months 24โ26
- Publishes the meeting calendar: what meets, how often, who decides, what document comes first.
- Writes the decision table and runs
decision_rights.py; it finds two decisions nobody would carry out.
- Moves gate reviews to a roster of six reviewers; the tooling enforces two approvals.
- A threshold change slips into an evaluation config; the cadence catches it in four days (section 10).
2. Why forty people need a cadence
- At three people, everyone knew everything.
- Decisions happened in conversation; the gate had two reviewers who sat together every day.
- At forty, information stops flowing by itself.
- Six teams, several time zones, and a CEO who is no longer in most conversations.
- Without a rhythm, decisions are made twice, or not at all, and the loudest person sets priorities.
- A cadence is a promise about time.
- Each question has a place and a time where it is decided, and a document that comes before the meeting.
- People can then work without asking permission, because they know when the next check is.
3. The rhythm, from daily to yearly
The operating rhythm (scenario)
Each level answers a different question at a different speed.
- Every meeting has three things written down.
- The decision it exists to make.
- Who decides.
- The document that must be read before it starts. No document, no meeting.
- Amazon's example.
- Amazon's practice of starting meetings by silently reading a six-page narrative, described by former executives in "Working Backwards", is the model for the weekly review and quarterly planning.
4. Quarterly goals tied to the one-pager
- Objectives and key results (OKRs).
- An objective says what matters; three to five key results say how you will know it happened, with numbers.
- The method comes from Andy Grove at Intel and was popularised by John Doerr. Doerr suggests ambitious key results where reaching about 70% counts as success.
- Gopala's rules for using them.
- Every company objective maps to a line on the strategy one-pager (S2 Part 3).
- At most three company objectives a quarter.
- Key results are outcomes, not activity ("8 new regulated customers", not "200 sales calls").
- Goals are not used for pay or performance ratings, so people do not set easy ones.
- Month 24โ27 company goals (scenario).
- O1 ยท Win the regulated beachhead: 6 new finance or health customers; NRR of the first cohort above 120%; one published reference.
- O2 ยท Start the environments moat: 1,500 runnable environments; a measured gain on held-out tasks from the first batch.
- O3 ยท Keep the gate independent at scale: zero updates merged without two recorded approvals; gate review median under three days.
| Objective | Key result | Owner | End of quarter (scenario) |
|---|---|---|---|
| O1 Regulated beachhead | 6 new finance/health customers | Head of sales | 5 (0.8) |
| O1 | Cohort-1 NRR > 120% | Customer success | 131% (1.0) |
| O2 Environments | 1,500 runnable environments | Research lead | 1,120 (0.75) |
| O2 | Measured gain from batch 1 | Research lead | Small but outside the interval (1.0) |
| O3 Gate independence | 0 merges without two approvals | Safety lead | 0 after week 101 (1.0) |
| O3 | Gate median < 3 days | Safety lead | 3.4 days (0.6) |
5. The weekly operating review
- Input metrics, not only results.
- Results (revenue, churn) arrive late. Inputs (trials started, tasks accepted, environments built, gate queue) can be changed this week.
- Amazon distinguishes "controllable input metrics" from output metrics for the same reason, as described in "Working Backwards".
- How the hour runs.
- Ten minutes reading the metrics pack in silence.
- Each metric owner explains only what moved outside its normal range.
- Decisions and owners are written in the pack before the meeting ends.
One week's operating review pack (scenario, week 103)
The six inputs every lead can act on this week.
- What it replaced.
- A monthly status meeting where each lead presented slides about activity.
- The rule for the CEO.
- Gopala asks questions in the weekly review but does not reassign work there; changes go through the leads.
6. Who decides what, checked by a script
- A decision table, in RAPID letters.
- Bain & Company's RAPID: Recommend, Agree (can veto), Perform, Input, Decide. Exactly one person or body decides.
- It answers the question forty people ask most: "Can I just do this?"
- One-way and two-way doors.
- Jeff Bezos's 2016 Amazon shareholder letter separates decisions that are hard to reverse (Type 1) from reversible ones (Type 2), and argues that most decisions are Type 2 and should be made quickly by small groups.
- In Gopala's table, training runs over $1M, gate rules and customer data agreements are one-way doors; most product and engineering choices are two-way doors and are not in the table at all.
- The checks.
- Exactly one decider per decision.
- Nobody both recommends and decides a gate decision (no self-approval).
- The CEO is never the only decider on a gate matter without a safety veto.
- Every decision has someone who performs it.
"""decision_rights.py โ check the company's decision table for the usual faults.
Roles per decision use RAPID-style letters (Bain & Company): R recommend,
A agree (veto), P perform, I input, D decide. The table is the story's
SCENARIO; the checks are the point:
1. exactly one D per decision;
2. nobody holds both R and D on a gate decision (no self-approval);
3. the CEO is not the sole D on gate decisions (the gate stays independent);
4. every decision has someone who performs it.
"""
T = {
"ship a model update (gate)": {"research lead": "R", "safety lead": "D", "gate reviewer 2": "A", "CEO": "I", "infra head": "P"},
"start a training run > $1M": {"research lead": "R", "CEO": "D", "board": "A", "infra head": "P", "finance": "I"},
"hire a team lead": {"CEO": "D", "team leads": "I", "people ops": "P"},
"price change": {"head of sales": "R", "CEO": "D", "finance": "I", "product head": "I"},
"new customer data agreement": {"head of sales": "R", "counsel": "A", "CEO": "D", "security": "P"},
"release to a new customer tier":{"product head": "R", "safety lead": "A", "CEO": "D", "infra head": "P"},
"change the gate's thresholds": {"safety lead": "R", "CEO": "D", "board": "A", "research lead": "I"},
"pause the self-improvement loop": {"safety lead": "D", "on-call engineer": "P", "CEO": "I"},
}
def check(table):
issues = []
for dec, roles in table.items():
ds = [p for p, r in roles.items() if "D" in r]
if len(ds) != 1: issues.append(f"{dec}: {len(ds)} deciders")
if "P" not in "".join(roles.values()): issues.append(f"{dec}: nobody performs it")
if "gate" in dec or "loop" in dec:
if any("R" in r and "D" in r for r in roles.values()): issues.append(f"{dec}: self-approval")
if ds == ["CEO"] and "A" not in "".join(roles.values()): issues.append(f"{dec}: CEO alone decides a gate matter")
return issues
def main():
print(f"{len(T)} decisions checked")
for i in check(T) or ["no faults"]:
print(" ", i)
bad = dict(T); bad["ship a model update (gate)"] = {"research lead": "RD", "CEO": "I", "infra head": "P"}
bad["change the gate's thresholds"] = {"CEO": "RD", "infra head": "P"}
print("a tempting shortcut (research lead ships its own updates, CEO tunes the gate):")
for i in check(bad):
print(" ", i)
if __name__ == "__main__":
main()- What the first run found.
- Two real gaps in the table as written: nobody was named to carry out a price change (billing must be updated) or a change to the gate's thresholds (the evaluation config must be edited and tested). Both were fixed by adding a performer: finance and the evaluation engineer.
- The tempting shortcut, letting research ship its own updates and the CEO tune the gate alone, produces three gate faults. That is exactly what Season 1 avoided with a second reviewer, and what section 10 shows happening by accident.
One-way doors in the decision table (scenario)
Who holds the D, and who holds a veto, for the decisions that are hard to undo.
7. Try it: break the gate
Decision rights for the gate (same rules as decision_rights.py)
- Things to try.
- Make the research lead decide on shipping: self-approval.
- Untick the board's agreement on thresholds: the CEO alone can loosen the gate, the risk the Series A clause was written to prevent (S2 Part 1).
8. The gate at forty people
- In Season 1, Gopala and one reviewer approved every update.
- Now research proposes updates every week, and two people cannot review them all without becoming the bottleneck.
- The roster.
- Six trained reviewers: three from safety, the research lead, the head of infrastructure and Gopala.
- Each update needs two reviewers, at least one from safety.
- Nobody reviews an update they wrote, or one written by someone they manage.
- Load is balanced, and Gopala is used last, so he stays a reviewer of last resort, not the default.
- The tooling enforces it.
- The model registry refuses to promote an update without two recorded approvals from eligible reviewers.
"""gate_roster.py โ assign two reviewers to each gate review without conflicts.
As the team grows, the gate (Part 11) needs more reviewers than Gopala.
Rules: a reviewer may not have authored the update or manage its author,
and at least one reviewer per review comes from the safety team.
Updates and reviewers are the story's SCENARIO; the assignment is greedy
and balances load, then reports any review it could not staff.
"""
REVIEWERS = {"safety lead": ("safety", None), "red-team eng": ("safety", "safety lead"),
"eval eng": ("safety", "safety lead"), "research lead": ("research", None),
"infra head": ("infra", None), "Gopala": ("exec", None)}
MANAGES = {"research lead": {"env eng", "data eng", "rl eng"}, "safety lead": {"red-team eng", "eval eng"}}
UPDATES = [("u-41", "env eng"), ("u-42", "rl eng"), ("u-43", "data eng"), ("u-44", "red-team eng"),
("u-45", "research lead"), ("u-46", "env eng"), ("u-47", "rl eng"), ("u-48", "eval eng")]
def ok(rev, author):
return rev != author and author not in MANAGES.get(rev, set())
def main():
load = {r: 0 for r in REVIEWERS}
for uid, author in UPDATES:
safety = sorted((r for r, (t, _) in REVIEWERS.items() if t == "safety" and ok(r, author)), key=lambda r: load[r])
if not safety:
print(f"{uid} by {author:13}: CANNOT STAFF (no eligible safety reviewer) -> escalate"); continue
first = safety[0]
rest = sorted((r for r in REVIEWERS if r != first and ok(r, author)), key=lambda r: (load[r], r == "Gopala"))
second = rest[0]
load[first] += 1; load[second] += 1
print(f"{uid} by {author:13}: {first} + {second}")
print("load:", ", ".join(f"{r} {n}" for r, n in load.items()))
if __name__ == "__main__":
main()How one update moves through the gate now
The same gate as Season 1, Part 11, with more reviewers and the rules in software.
- What the run shows.
- Every update got two eligible reviewers; the red-team engineer carried the most (four), a sign that safety needs its next hire soon.
- Update u-44, written by the red-team engineer, could not go to the safety lead, who manages its author.
- Gopala reviewed two of eight updates instead of all of them.
9. The board's role
- Quarterly meetings, monthly updates.
- The monthly update is one page: cash, burn against plan, ARR, hires against plan, the gate report, the top three risks.
- The quarterly meeting reviews strategy, approves the budget and one-way-door decisions, and spends time alone with the independent director.
- The gate report.
- Every quarter the safety lead presents the gate log to the board directly: updates proposed, approved and rejected, with reasons, any merge exceptions and any loop pause.
- Gopala attends but does not present it.
- What the board does not do.
- Choose research directions or product features. It decides whether the plan, the money and the risks fit together.
10. The week the cadence was tested (scenario)
- Week 104: a research engineer, trying to unblock an update that kept failing, changes a regression tolerance in an evaluation config from 1.0 to 2.5 points as part of an unrelated cleanup. The change passes code review by a teammate, who does not notice that the file is part of the gate.
- How it was caught.
- The weekly operating review showed gate approvals jumping from 3 to 7 in a week with no change in the research plan: an input metric outside its normal range.
- The safety lead traced it in a day to the config change.
- Total time from change to rollback: four days. Two updates approved under the looser tolerance were re-reviewed; one was rejected.
- What changed afterwards.
- Gate configuration files now live in a separate repository owned by safety, with the decision table's rule built in: only the safety lead can merge, and threshold changes need the board's recorded agreement.
- A test fails the build if any gate threshold differs from the approved values.
- The engineer was not blamed: the post-incident review found a system that let a reasonable cleanup touch the gate.
- Why this is a success.
- Season 1's gate relied on two people paying attention. Season 2's relies on a metric, a roster, a separate repository and a test.
11. A meeting budget
- Meetings grow faster than headcount if nobody stops them.
- Gopala sets a budget: no more than about a fifth of an engineer's week in recurring meetings.
- Writing replaces most status meetings.
- Weekly written updates per team, read before the operating review.
- Decisions recorded in a decision log with the RAPID roles, readable by everyone.
Recurring meeting hours per week by role (scenario)
Story estimates; engineers stay near 10% of a 40-hour week.
12. Leading signals the cadence watches
| Signal | Why it leads | Where it is reviewed |
|---|---|---|
| Trials started and activated | Revenue three months later | Weekly review |
| Accepted-change rate per customer | Churn and expansion | Weekly review |
| Gate queue and median review time | Release speed, and pressure to loosen the gate | Weekly review, gate report |
| Environments built and their pass rates | The Phase 6 moat | Weekly review |
| Hiring pipeline vs plan | Burn and delivery next quarter | Monthly review |
| Merges without two approvals | Gate integrity | Weekly review (target 0) |
13. What a decision document looks like
- Every one-way-door decision starts as a document of at most six pages.
- Written in sentences, not bullet fragments, so the reasoning has to connect.
- Read in silence at the start of the meeting.
- The template.
- The decision asked for, in one sentence, and who decides (from the decision table).
- Context: what has changed and why now.
- Options considered, including doing nothing, with costs and risks.
- The recommendation and the evidence for it: the script, the evaluation or the customer data.
- What would make us reverse it, and when we will check.
- Appendix: the numbers, with sources and labels (measured, simulated, scenario).
- An example (scenario).
- "Start Maestro-2 small-scale predictions in month 27" ran four pages, with the scaling-law plan from Season 1, Part 2 in the appendix. The board approved it with one condition: a checkpoint review at month 29.
14. Quarterly planning in one day per team
- Before the day.
- Each team writes a two-page plan: last quarter's goals scored honestly, what it learned, its proposed goals, the people and compute it needs.
- The day.
- Morning: each lead presents in 20 minutes; others ask questions.
- Afternoon: the leads and Gopala resolve conflicts: two teams needing the same people, compute or customer.
- Evening: the company goals are written and published the next morning.
- A capacity rule.
- Roughly 70% of research and engineering time on the current plan, 20% on improving tools and paying down debt, 10% on exploration, so the research engine of Phase 6 has room before it has its own budget.
- The split is our assumption, a common rule of thumb rather than a measured optimum.
Where engineering time goes each quarter (our rule of thumb)
Reviewed every quarter; Phase 6 gives research its own budget.
15. How the CEO gets feedback
- The higher the title, the less honest the feedback.
- Gopala asks for it on a schedule instead of hoping for it.
- Three channels.
- A twice-yearly 360 review run by the independent director: every lead and a sample of the company, anonymous, with the summary shared with the whole board.
- A monthly skip-level lunch with five people from different teams.
- An executive coach, outside the reporting line.
- What the first 360 said (scenario).
- Strong on clarity and on the gate; weak on letting go of research decisions; too many late-night messages.
- His response: research direction is set at quarterly planning, not in chat; messages after 8 p.m. local time are scheduled for the morning.
16. Eight operating mistakes
- Meetings without a decision to make or a document to read first.
- Goals that measure activity instead of outcomes.
- Using goals for pay, so everyone sets easy ones.
- A decision table nobody checks for missing owners.
- Letting the people who propose updates approve them because the gate is slow.
- Keeping gate rules in the same repository as research code.
- A CEO who reassigns work in the weekly review, over the leads' heads.
- Blaming the person instead of the system after an incident.
17. Delegation box
- The rhythm, the company goals and the decision table
- The board relationship
- The gate's rules (the safety lead runs it and reports to the board)
- Assembling the weekly metrics pack
- Running
decision_rights.pyandgate_roster.pyon every change - Drafting meeting notes and decision-log entries for owners to confirm
- An executive coach for the first year as CEO
- Board governance advice from counsel
- Tools: the decision log, the model registry, the metrics pack
18. Budget and risk tracker
| Item | Status after Part 4 | Note |
|---|---|---|
| Cadence | Published: daily to yearly | Every meeting has a decider and a document |
| Gate | 6 reviewers, 2 approvals enforced | From gate_roster.py |
| Decision table | 8 one-way-door decisions, 0 faults after fixes | From decision_rights.py |
| Gate risk | Lower | Separate repository, threshold test |
| Speed risk | Watch | Gate median 3.4 days against a 3-day goal |
| Safety capacity | Watch | Red-team reviewer carries the most load |
19. The series
The Maestro, Season 2: 23 parts in five phases (Season 1: 18 parts)
20. Reading the numbers
- Primary / standard
- Bain's RAPID; the 2016 Amazon shareholder letter; Doerr's OKR method; "Working Backwards" on narratives and input metrics.
- Secondary โ ๏ธ
- None in this part.
- Our assumptions
- The decision table, people, updates, goals, metrics, meeting hours and the incident are the story's scenario.