Chapter 9The Consensus Cycle: How Unified Context Is Created

1. First confirm that everyone is facing the same set of facts #

A company's top can see a threat early, and can keep pressing the team to move faster, and still not know what the company can actually do. Research from Nokia's turn into smartphones shows how that happens.

In 2016, Timo Vuori and Quy Huy published a qualitative study of Nokia in Administrative Science Quarterly. They used 76 interviews to look back at the smartphone business from 2005 to 2010. The people interviewed included senior and middle managers, engineers, and outside experts. The 76 is a count of interviews, not of 76 distinct people. What the study sets out to explain is also not Nokia's whole rise and fall. It is how organizational structure and shared emotion affect the flow of information and judgments about innovation (Vuori & Huy, 2016).

The study summarizes the shared emotion of the top and the middle as fear pointed at different objects. Attention at the top pointed at outside competitors such as Apple, and at pressure from shareholders and the market. Attention in the middle pointed more at superiors' reactions, peer competition, resources, and jobs. Under pressure, the top asked for faster delivery, did not pass the full severity of the outside threat to the middle, and also read reports in the light of its own hopes. The middle reduced the bad news it sent up, and sometimes gave a more optimistic account of progress and capability than the facts supported (Vuori & Huy, 2016).

The authors later recalled, in an ASQ interview, that one middle manager described his superior as a "filter" sitting between him and the very top, and that one senior said information had already been beautified by the time it reached him. The two sides were not wholly without information. They were not facing the same version of the key facts together. The top knew the competitive threat, and overestimated how soon the organization could deliver. The middle sat closer to the limits of the software and the projects, and did not send those limits, safely and steadily, to the people who had to decide.

That interaction left the top with an overly optimistic picture of technical capability, and also pulled more of the organization's attention onto short-term product delivery, so that longer-term capability was harder to resource. The researchers call the latter temporal myopia: the next release grows tighter, and long-term investment grows harder to fund. Higher delivery pressure did not fill in the facts. It raised the incentive to report optimistically.

What the study identifies is an organizational mechanism: when the personal cost of reporting bad news is higher than the cost of keeping it, different layers decide on different versions of reality. It cannot treat "fake context" as the only cause of Nokia's loss. The rapid success of Apple and Android was already an important outside change. Symbian's technical debt, software capability, platform ecosystem, touch experience, product-line coordination, market timing, and later strategic choices also cannot be ruled out from this single-case study. The authors also discussed how software expertise at the top affected the process.

Nokia shows that a shared standard cannot repair facts that are distorted or incomplete. Even if everyone agrees on "ship a competitive smartphone as soon as we can," the sentence still cannot support a high-quality decision if the people responsible for different steps do not know what the others hold, which progress is still a hypothesis, and which risks have not been reported.

Unified context is not the same as "communicate more." The people in the same class of decision need to be able to check: whether they are using the same version of the facts; which materials have a source, and which are only estimates; what is still unknown; who is responsible for which stretch of information; and what new evidence would require the decision to be looked at again. Answers may differ. The disagreement has to be locatable in the standard or in the facts.

Chapter 6 shows how a judgment is written as a standard. Chapter 7 handles where a standard comes from. Chapter 8 handles how a true judgment is seen and understood by someone else. Only after those moves cross a team boundary can they support company-wide shared judgment and dispersed action.

2. One cycle, four ranges of use #

Chapter 6's standards cycle has seven moves:

Judgment → Distill → Form the standard → Issue → Execute → Return → Iterate.

A decision first leaves the standard and the context that were used, then distills a reusable relation from them. The candidate judgment is written as a standard with conditions of use, evidence requirements, and a stop line, and enters the actual work. Results, exceptions, and risks then return to the owner, and push the next revision. The team does not have to guess from scratch each time, and it also does not treat a first version, once issued, as an order that is forever correct.

The Clarity Method supplies the cycle's entrance. After a team finds that an angle was not looked at, that a scale of judgment is missing, or that the conditions of use have not been confirmed, it looks first for a candidate standard already verified outside, checks whether it can migrate to the problem at hand, verifies it with a cheap action, and then writes the result back into the standards cycle. It is not another process. It is one method by which an organization finds and verifies a standard.

The Interference Method supplies understanding and adoption. That a standard is written correctly does not mean the person using it has already seen the relevant facts, understood the conditions of use, or known how the standard relates to the decision in their hands. The organization first has to understand what members already hold, and then, in a scene where a decision is actually required, state information that can be checked, is complete, and can be delivered, and let questions and counterexamples enter the discussion.

When the Interference Method is used inside an organization, differences of power push the ethical boundary to the front. A manager cannot pick only the true information that favors them, and wrap a pre-decided answer as "consensus." Silence from staff also cannot be read as agreement. Treat People with Honesty, established in Chapter 8, still holds here: the claim has to be verifiable; material facts that would change the other party's decision cannot be hidden; promises must be deliverable; and the other party's capacity for informed judgment and for dissent has to be kept. What internal Interference Method narrows is misunderstanding, irrelevant information, and options that cannot be delivered. It does not narrow bad news, disagreement, or channels of appeal.

These are not three parallel paths. They are the same cycle enlarged layer by layer. The moves of the cycle do not change: find a gap in judgment, form and verify a standard, let the relevant people obtain the necessary facts, act inside a stated boundary, and let results come back to revise.

Figure: Part III shares this one main figure. The Clarity Method helps find and verify a standard; the Interference Method helps a true judgment be understood and adopted. What changes is the cycle's range of use, not a second system.

In a person, this set of moves is called adaptive insight. A person finds which standard they are missing, borrows from outside, checks the conditions of migration, and then revises against reality.

In a team, it is called the standards cycle. The judgment is written down. Other people and AI can check it, use it, and send results back, and it can keep being revised when the original author is not in the room.

Across departments and sites, when the people in the same class of decision begin to share the judgment standard, the facts that belong to the decision, the disagreements already known, and the responsibility for action, it becomes the company-wide consensus cycle. It is not copying a slogan to more people. It is letting different decision-makers work independently around the same set of material that can be checked.

When people outside the organization, without taking an order from it, begin on their own to adopt the language and the judgment standard, and use it to name a problem, compare plans, or constrain their own action, that use is called cultural spillover. It is still the same set of judgment moves traveling farther. It is not a fourth system of management.

The consensus cycle cannot replace strategy, technology, capital, governance, and delivery, and it cannot guarantee that the organization is forever right. What it supplies is a condition: after a judgment leaves a few heads, it can still be seen, questioned, executed, and revised.

3. Consensus is not believing the same answer #

"Consensus" is often used to mean that everyone accepts the same conclusion. On that definition the organization has two remaining moves: keep meeting until the objectors are persuaded, or have the very top give the answer and call the absence of open opposition agreement. Neither move can support the dispersed judgment an AI-native organization needs.

Consensus is a shared state that can be checked:

The people in the same decision can use the same set of judgment standards and the same version of the facts that belong to the decision, and they know which disagreements and unknowns remain. After the decision is made, they also know who holds the decision right, and what action each person has committed to take.

The definition does not require everyone to believe exactly the same conclusion, and it does not require every decision to be voted. It requires that disagreement can be located, that the decision can be executed, and that results can come back.

Suppose a company has to decide whether to ship a new refund feature. The product lead proposes shipping it to cut user wait. The support lead submits a record of new kinds of fraud. The finance lead lists the cost of refunds. That the three claims differ does not mean the organization has not yet formed consensus.

They can first confirm a shared judgment standard: whether the user's problem is actually solved, whether the fraud risk sits inside a bearable range, and whether the added cost meets a return written in advance. They can then confirm the facts that belong to the decision: present wait time, the fraud sample, refund amounts, the policy version, and experimental results. They also leave the disagreements already known: they still read differently whether the fraud sample represents the full risk, and data in the next two weeks may change the conclusion.

If the person accountable for this class of shipping decision has already been named, that person can, on the shared material, choose to open the feature first to a small set of users, with an amount cap and a stop condition. The other participants do not have to convert to one belief. They do have to confirm their action commitments: how support will mark anomalies, when finance will supply cost data, and on what signal product will stop the trial. Two weeks later, the results return, as agreed, into the standard and the context.

That is the difference between consensus and unanimous opinion. Shared material makes different judgments comparable. A named decision right lets the organization not wait for a unanimous vote. Action commitments keep the decision from stopping in the minutes. Known disagreements and revision conditions leave an opinion that was not adopted a chance to be opened again by new evidence.

Consensus is also not obedience. A member can say plainly "I do not accept this reading of the risk," and still understand the present standard, know who made the decision by law or by duty, and commit to act by the decision until the next review. Conversely, a person who says they agree, and then keeps using another version of the data and another ranking of priorities, has not formed consensus either.

Shared standards and shared context are therefore not a reason for centralized decision. They are the public input of dispersed decision. Different regions can make different choices for different customers. Different product lines can use different methods inside the same risk boundary. As long as the basis of the trade-off, the source of the facts, the permission, and the result can be traced, those differences are adaptation to the scene, not a loss of organizational control.

A one-person rule is exactly what lacks those conditions. If the boss's conclusion cannot be rebuilt as a standard and as facts, if dissent cannot enter the record safely, and if every decision still has to return to the boss in person, what the organization has is concentrated control. Acceptance of the consensus cycle is not "whether people look more like the boss." It is "when the boss is not in the room, can the right person make an explainable decision on the shared material; and when new evidence appears, can the original standard be revised."

4. Unified context unifies only the facts that belong to the decision #

The part of the definition of consensus that is easiest to enlarge is "the same version of the facts." If that is read as everyone being able to see all of the company's information, unified context quickly becomes a transparency project that cannot be finished, and should not be finished.

Pay, medical information, unannounced mergers, customer privacy, security incidents, and regulated data all have lawful permission boundaries. Different roles also do not need to read every file that has nothing to do with their decision. A support person handling an ordinary refund does not need to see an acquisition plan. A designer does not need all customer identity data merely because they sit in a product review.

The object of unified context is not the whole company. It is one decision, or one class of repeated decisions. The test of acceptance is this: the people accountable for that class of decision can, inside their permissions, obtain the same version of the necessary facts, and they know which information is still unknown, restricted, or in dispute.

A usable context has to answer at least a few kinds of question:

1. What are the facts. Present data, customer feedback, constraints, and historical change, each separately.
2. Where the facts come from. Whether source, method of collection, time range, and owner are clear.
3. Which version is in use now. When the policy, the data, and the assumptions were updated, and whether the old version has already left the work.
4. What is still unconfirmed. Which contents are only estimates, which evidence conflicts, and which information is temporarily not visible.
5. Who needs these facts in order to decide. Whether permission and decision responsibility match.

Putting a large number of files into a shared space does not answer those questions. Ten conflicting policies that can all be searched only show that there is a lot of information. Meeting notes with no version and no owner only show that someone once recorded something. A document that everyone can see, if it has not entered the actual entrance of the work, will still leave executors using the older version that is more convenient in their hands.

Unified context is also not a unified interpretation. Two people looking at the same fraud data can make different forecasts of where the risk is going. One may say the sample is enough to pause; the other may ask for a wider watch. The disagreement then sits clearly in the inference and in risk preference. What has to be handled is the standard or the assumption, not a further general demand to "communicate more."

When permission is insufficient, the organization has two honest moves. The first is to supply, inside the lawful range, a summary that is enough to decide, and to mark clearly what has been hidden and how it might affect the choice. The second is to move the decision right to someone who can obtain the necessary information. The most dangerous move is to make a person responsible for the result, withhold the facts the decision requires, and then read a wrong guess as a lack of capability.

Context also changes faster than a standard. The same refund standard may be used for a year, while the present method of fraud, the policy version, and the customer record update every day. The organization therefore cannot name an owner only for the standard. It also has to name, for the key facts, a source, an update frequency, access permission, and a condition of expiry. Otherwise a unified standard will copy expired facts more stably to more people and to AI.

5. Context, not Control: put the decision at the scene #

On April 24, 2017, Source Code Capital published Zhang Yiming's talk from that year's Code Meeting, Being a CEO Means Avoiding Rational Conceit. The talk uses Context and Control to discuss management after a company has grown.

Netflix's culture deck, circulating publicly since 2009, had already listed "Context, not Control" as a management principle, and had explained Context with strategy, metrics, assumptions, goals, roles, and transparency of decisions (Netflix, 2009). Zhang Yiming's talk then left first-hand material on how a Chinese internet company understood and practiced the principle.

Zhang defined Context as "the set of information a decision requires," including the rationale, the market environment, the industry structure, priorities, how far the work has to go, and business and financial data. Control includes committees, directives, decomposition and roll-up, process, and approval. What he argued for was not the cancellation of management. It was "ample Context, little Control": let more of the people who hold the facts on the ground take part in the decision, and have managers intervene only when necessary (Source Code Capital, 2017).

The claim first handles a bottleneck of decision. When a company is small, the CEO may know the business best, and can also handle most of the important choices in person. After the business expands, the industry grows more complex, the CEO's time is split among fundraising, hiring, public affairs, and organizational management, and many people on the line hold more complete concrete information. If every file has to be rolled up layer by layer to the CEO, and then wait for the CEO's instruction, the organization both loses the information on the ground and locks its speed of response to one person's calendar.

Putting Context at the scene is not a way of letting everyone make every decision. It is a way of letting the people who hold the matching role and permission act without having to guess what a superior intends. Management has to state aims and priorities, publish the business information that belongs to the decision, and explain past failures and present constraints. Members then have to propose plans on that basis, own the result, and bring outside change back into the organization.

The talk offered one management reflex: when a piece of progress goes wrong, ask first whether Context is insufficient, rather than adding Control at once. Managers need to check whether industry change, business data, and past failures have already been shared. They also need to tell whether a better judgment came from capability, or from information other people cannot reach. Zhang also noted that some managers "use information asymmetry to demonstrate their own value." When information stops in one person's hands, that person becomes a required node for every decision. The larger the organization, the longer everyone else waits.

The practices Toutiao described at the time sat with that principle. Management OKRs were open to people below; members could see the aims of superiors, peers, and other departments, and know what the company was ranking first. The company encouraged group chat and cross-department communication, and also had the CEO answer employees' questions in person. Zhang also said the company had an internal tools team of nearly a hundred people, had built its own OKR system and wired it to internal instant messaging, and that a new hire could see internal material and understand collaboration duties relatively quickly.

Those practices show how context obtains a carrier: aims can be looked up; relevant discussion does not stop only in one-to-one talk; an internal system lets obtaining information depend less on whom one already knows. Tools only lower the cost of finding, sharing, and maintaining. They do not, by themselves, judge which facts matter, and they do not guarantee that the material is true, current, and free of a material omission. If three versions of a policy sit in one system, faster search only lets the error arrive faster.

The principle also has stated exceptions. In the talk Zhang kept several situations that need more Control: a major crisis, and a key project whose time window is very short; a new business, a new department, or a new senior hire that has not yet obtained the necessary context; and a clear mismatch between the person and the role, which needs a temporary intervention. Context, not Control, was therefore never "no control." It requires the manager first to show why control is necessary, and then to reduce intervention as members become able to carry the decision.

The Nokia study shows what follows when information from above and below cannot fully meet: the top holds the outside threat, the middle knows internal capability better, and fear keeps the two kinds of fact from joining reliably. The mechanism Zhang described hands decision background, aims, and business data to more people on the ground, and reduces the dependence on information traveling only up and down the hierarchy.

The two sets of material show a direction of mechanism. They do not constitute a comparison of operating wins and losses. The talk is a first-hand full text; the internal tools and practices remain a company account. The public material does not give decision time, rework, reporting of bad news, speed of correction, or operating results before and after those practices went live, and it cannot prove that ByteDance has already achieved unified context. Context, not Control is a management mechanism. It is not proof that unified context has already been reached.

6. Make the consensus cycle actually happen #

A talk, a collaboration tool, or an all-hands announcement cannot make the cycle run on its own. The consensus cycle has to join the material a decision leaves to the next action and the next revision.

Its smallest unit is not a meeting, and not a document. It is a consensus record that a later person can keep using. For an important or repeating decision, that record contains at least six parts:

1. The judgment standard. Which value ranks first, what evidence counts, and which boundaries cannot be crossed.
2. The facts that belong to the decision. The data, materials, sources, versions, constraints, and assumptions now in use.
3. Known disagreements and unknowns. Which facts are still in dispute, which inferences have not been verified, and who holds a different view.
4. Decision rights and permissions. Who is responsible for which layer of decision, and what situation must escalate.
5. The present decision and the action commitments. What was chosen, and what each participant is responsible for next.
6. Conditions of revision. What signal would overturn the present judgment, when it will be reviewed, and who publishes the next version.

Those six items let Chapters 6 through 8 meet in one place. The standards cycle supplies the judgment standard, the version, and the return. The Clarity Method helps find the gap, look for a candidate standard, and verify it. The Interference Method helps different roles understand the true information and the limits the decision rests on. The consensus cycle then fills in the facts, the disagreements, the permissions, and the commitments, so that the material can be used across teams.

A meeting has several legitimate outputs here, and does not have to invent a new standard every time. It can make a decision on an existing standard and assign commitments. It can find that a stretch of context is expired and require the owner to correct it. It can propose a candidate revision or a hypothesis still to be tested. It can also record, plainly, the disagreements not yet resolved, the missing evidence, and the next owner. A meeting that only synchronizes information is not, for that reason, invalid, provided the participants actually need that information to complete the matching decision.

A document also does not form consensus because it has been written down. It has to be findable by the people who decide, and accessible inside their permissions. Source, version, owner, and effective time have to be clear. The old version has to leave the actual entrance. Users also have to be able to point to where it does not apply and where facts are missing. If, when the original author is not in the room, other people can still only guess what was meant, the document is only a backup of personal memory.

A review that actually updates a standard has to walk an observable path. The team first sets the result beside the original standard, the assumptions, and the stop condition, then judges whether the present standard should be kept, revised, or withdrawn. The owner states what changed, why it changed, and when it takes effect. The new version enters the workflow and the AI task; the old entrance is closed. The next decision of the same kind then checks whether members actually used the new version. Summarizing only that "communication was insufficient this time" or that "we will pay more attention later," without changing any field or any entrance, cannot count as a standard already updated.

Bad news has a special place in this cycle, because it is the material most likely to disappear before it reaches the person accountable. The organization needs a stated rule of non-retaliation, an independent path of escalation, and a handling deadline, and it also has to show publicly how a problem was later handled. Calling for "candor" while every report stops in a private chat will make the reporting channel fail. Only when the person accountable states, in public, which standard was revised on the basis of the bad news does reporting and revision acquire a visible link.

A consensus record also keeps the Interference Method from crossing a line inside the organization. The person who proposes a plan has to leave the unfavorable facts and the limits of use at the same time. The person who decides has to see the objections still unresolved. The person who executes knows that what they have committed to is an action, not a rewrite of their inner judgment. The organization can then decide inside a limited time without having to process dissent into agreement.

Transn's public material supplies one entrance to a rebuild: it set a CAIO and an AI-native decision committee, required AI projects to supply a runnable demo, and used internal use and satisfaction to form an "energy gold" incentive. The material comes mainly from company event reports and promotional articles. The effect has no third-party verification. A demo is only one kind of evidence; that it runs does not mean it solved the right problem. Usage and satisfaction likewise cannot, on their own, stand for business value. Transn's full rebuild is in Chapter 14.

7. Humans are responsible for consensus; what work AI takes over #

"Humans are responsible for consensus" does not mean humans only discuss direction and AI is responsible for everything else. That split writes a changing technical capacity as a completed fact, and it also ignores physical execution, legal duty, safety risk, and many new problems that have not yet formed a standard.

The more accurate division is this: humans are responsible for consensus and for final accountability; AI takes the rest of the work that can be standardized, automated, accepted by a result, and kept inside a controllable risk.

"Responsible" does not mean a person has to complete every step of forming consensus by hand. AI can help retrieve related material, mark sources and update times, compare policy versions, find conflicts in the data, arrange the evidence of different plans, compile the disagreements already known, check whether a decision violates a live standard, and remind the owner when a review is due. As long as the output can be checked, that work can cut the cost of maintaining context by a large amount.

Once standards and context are clear, AI can also complete a large volume of execution: generate copy and code, handle reports, answer common questions, call systems, watch for anomalies, and hand a task back to a person on the escalation condition. Each human revision and each business result then enters the record, and becomes material for the next version of the standard and the context.

Humans still have to carry several kinds of responsibility that do not disappear because a model is performing well. First, decide for whom the organization creates what value, which item ranks first when they conflict, and which boundaries cannot be traded. Second, judge which facts are important enough for the present decision, and which risks must be exposed. Third, allocate decision rights, data permissions, and paths of escalation. Fourth, make a commitment and carry the consequence when standards conflict or evidence is insufficient. Fifth, approve the keeping, revision, or withdrawal of a standard.

Whether a piece of work is fit to automate has to check at least four things: whether the judgment standard can be stated; whether the necessary context can be supplied lawfully and in time; whether the result has observable evidence of acceptance; and whether an error can be found, stopped, or recovered inside a cost that can be borne. Medical, financial, safety, employment, and other regulated decisions still have to obey the matching professional duty and the law. When the conditions are not complete, AI can assist investigation and propose a candidate. It cannot obtain the final decision right merely because execution is cheap.

The Air Canada case has its main ground in Chapter 2. The echo here is only this: when an automated front door and a formal policy used conflicting context, the organization did not lose its responsibility because the answer was given by a tool. The decision did not establish whether the tool used generative AI, and it is therefore not evidence of "hallucination" (Moffatt v. Air Canada, 2024 BCCRT 149).

8. Cultural spillover: people outside begin to use the language of judgment #

The consensus cycle's range of use can keep expanding outward, and it is not enough that people outside have heard of a company. One news story, a popular slogan, or high search volume can only show that attention appeared. Cultural spillover appears when people outside begin, on their own, to use the judgment standard and the language to name a problem, compare plans, record a decision, or constrain action.

Crossing the Chasm is a short example. Geoffrey Moore published the book in 1991, and named and systematized a break that often appears when a high-tech product moves from early adopters toward the mainstream market. The five groups of adopters were not invented by Moore. They come from Rogers's earlier work on the diffusion of innovations. Moore's work was to put the weight on that break, and to supply a language in which it can be discussed in common (Rogers, 1962; Moore, 1991).

In the author's note to the 2014 third edition, Moore recorded that the book would pass among colleagues and spread a shared vocabulary. In 2000, an independent report in Strategy+Business also recorded that the set of words had entered the daily language of Silicon Valley executives; one software-company CEO summarized its value as letting people who had just met discuss the same class of market problem more quickly (Moore, 2014; Fisher, 2000). Those materials can support "the language was adopted outside on its own." They do not give a clear denominator, and they cannot prove that every user follows the same judgment standard.

Cultural spillover therefore has three boundaries.

First, it is not the same as brand. Brand includes identification, association, trust, preference, and economic value. It does not measure attention alone. Awareness is what mainly answers whether something has been seen. Cultural spillover asks a narrower question: have people outside the organization brought your language of judgment into their own decisions.

Second, it is not a proof of success. A wrong, expired, or even harmful judgment can also spread widely. That a word is used only proves that transmission happened. To judge whether it creates value, one still has to look at how users act, what the results are, and whether they revise after a counterexample appears.

Third, it cannot be declared by a communications project. A company can promote a word. It cannot complete, for outside users, the act of adopting it on their own. Observable signals include: an outside team writing the standard into its own decision record; customers using it to compare plans, and not only repeating a slogan; peers continuing to use, rewrite, and mark the conditions of use when there is no commercial partnership and no chain of command. A spillover rate usually has no reliable denominator, and should not be written as a precise percentage.

That also explains why cultural spillover is not a new parallel system. Inside the organization, a judgment first has a standard, the relevant facts, disagreement, and feedback. Outside users then take the same moves into their own scenes. The range of transmission has enlarged. The judgment still has to be tested against reality.

The Nokia, Zhang Yiming, and Moore materials carry, respectively, a mechanism, a management claim, and a sign of outside adoption. They cannot be used to compute the average effect of a tool, and they cannot guarantee that adopting this method will produce operating success. Unified standards and unified context supply a necessary condition for dispersed judgment. "Fill Context first, then add Control" is a management choice that can be practiced, checked, and revised. Cultural spillover only shows that people outside have begun to use a language of judgment.

Independent data on whether ByteDance's tools improved decision quality, a complete population of outside adoption, and a longitudinal rebuild sample from a traditional firm still have to be filled in. They help test the method's effect under different scales and constraints. They do not change the mechanism claim already narrowed here.

The standards cycle lets a judgment leave a person. The Clarity Method helps an organization find and verify a standard. The Interference Method helps a true judgment be understood and adopted. The consensus cycle then lets the standard, the context, the disagreement, and the responsibility for action be shared across teams. The more complete the shared material, the less the decision has to sit in a few people's hands.

Once standards and context are enough to support dispersed judgment, how many roles that only execute and only carry information does the organization still need? Which people can obtain a larger decision right, and how is it shown that they are already ready to carry the responsibility of judgment? Chapter 10 takes up that set of questions about the structure of talent.

What to Do Monday Morning (principal-leader view) #

Do not begin by auditing the company's "degree of consensus." Pick one decision that happens every week and that needs two or more teams, and run the first round on one page.

1. Name the decision and the person accountable. Write what the decision is, who makes it, which participants supply facts or carry action, and what situation must escalate.
2. Write the shared inputs. On the left, the judgment standard: ranking of value, requirements of evidence, and boundaries that cannot be crossed. On the right, the facts that belong to the decision: source, version, time of update, constraints, assumptions, and unknowns.
3. Leave the disagreements and the commitments. Do not require dissent to be rewritten as agreement. Record whether the disagreement sits in the standard, the facts, or the inference, and write the present decision and what each person is responsible for next.
4. Run a change-of-person test. Find two people who did not take part in the original discussion and who have the matching duty, and have them judge independently from this page. Record where they have to guess. Do not set a fixed threshold that has no verified basis.
5. Let the results come back. Use AI to compare versions and list conflicts and omissions. The owner decides whether the standard and the context should be kept, revised, or withdrawn, publishes the new version, and closes the old entrance.

When a subordinate or an AI brings back a wrong result, check first whether the context was sufficient, then the standard, the permission, the capability, and the execution. Context is not the answer to every problem. It is one class of cause that should be ruled out before adding an approval or replacing a person.

Chapter Acceptance Self-Check (against chapter acceptance standards) #

1. Claim restatable in one sentence ✓: the work least to be outsourced is shared standards, decision-relevant facts, known disagreements, and action commitments; humans own that consensus; AI takes work that can be standardized, automated, accepted, and kept inside controllable risk; outside use of the same language is cultural spillover.
2. Eight-section spine and whiteboard figure ✓: same facts → one cycle, four ranges → consensus is not one belief → unified context is not all information → Context, not Control as mechanism → consensus record / meeting / document / review → human–AI split → cultural spillover. Figure is one seven-move cycle enlarged by range; assembly, production lines, and workshops are out.
3. Evidence caliber ✓: Nokia from Vuori & Huy 2016, 76 interviews not 76 people, temporal myopia and filtered facts as mechanism, Apple / Android / Symbian / later strategy not excluded, "fake context beat Apple" out. Zhang Yiming from Source Code Capital 2017; Netflix 2009 for the phrase; Control exceptions kept; tools as company account, not proof ByteDance has unified context. Moore 1991 / 2014 and Fisher 2000 for outside language, not invention of the five groups. Air Canada as Chapter 2 echo, not a hallucination finding. Transn as a Chapter 14 preview, mechanism nouns only.
4. Case contract ✓: Nokia for filtered facts; Zhang for fill-Context-then-Control; Moore for outside naming; Duan, Musk, CCP, MD Anderson, and the three-museum spillover set are out.
5. Definitions ✓: consensus is not shared belief or a unanimous vote; unified context is one class of decision and one version of necessary facts; cultural spillover is outside use of the language of judgment, not brand, not a communications project, not a success proof.
6. Chapter boundary ✓: no engine / assembly / workshop metaphor; no "AI is responsible for everything else"; no 30% restatement or six-month stop thresholds; daily three questions left to Chapter 11.
7. Fluency ✓: rewritten in English voice from the Chinese authority; current prose-standard.