Updated November 2026: all twelve books are now finished, and every one has its full notes below.
Twelve books, ten lessons, eight early signals and one evidence test. Jump to: The four I would hand a board first · The early signals · The evidence test
The reading list: The Leadership Dozen. The next twelve: The Next Leadership Dozen.
Not sure which book to read first? See Where to start, by interest.
Ten lessons from twelve books
- Trust is a decision to accept exposure. It is not evidence that the exposure has gone away. Most of the failures on this shelf began with someone trusting a name in place of looking at the thing itself.
- Whoever controls the root controls the system, whatever the org chart says.
- Speed moves judgment upstream. When decisions outrun review, the only real human input is the guardrail written before the clock started.
- A score is only as honest as the person who set the weights. If an instrument has never embarrassed its sponsor, it is measuring loyalty.
- A dependency is a strategic choice even when nobody remembers making it.
- An exception without an expiry date becomes a policy without an author.
- A warning that has been true too many times without catastrophe stops sounding like a warning.
- The rewards concentrate and the costs disperse. That is why these schemes run for years, and why the early signal, not the courtroom, is the real protection.
- Question the expensive assumption, then let reality grade the answer.
- Character is what you do with power when the easier path pays. It is not shown by staying away from power.
The rest of these notes is where those lines come from.
The comfort of a green dashboard
There is a particular comfort in a green dashboard. The boxes line up. The numbers sit in range. Someone has done the work of turning a messy world into a surface a busy leader can read in thirty seconds. Sometimes that work is excellent. Sometimes the dashboard is the best organized part of a system nobody has looked at closely.
In September I published The Leadership Dozen, twelve books chosen on a simple idea. If artificial intelligence is moving from answering questions to taking actions, leaders need a more demanding reading diet than another stack of books about AI. Only one of the twelve is about AI on its face. The rest are about deception, war, finance, concentrated power, engineering and institutional change. All twelve are now finished.
One theme emerged, and it was not the one I expected. The worst institutional failures in these histories rarely began with no information. Almost always, someone had the information and it did not fit the story. An engineer had a damaged seal. A bank had an oddly structured payment. An analyst saw a model fail in the field. The organization also had a procedure, a credential or a reassuring report, and the contest between the uncomfortable fact and the comfortable account was rarely fair.
The deeper story is about people. Deception, ambition and collusion can produce extraordinary private rewards while the people who insist on the truth are overruled. Sometimes there is a conspiracy. More often it is ordinary people protecting careers, contracts or access. The costs seldom stop with the decision-makers. Employees absorb retaliation. Investors and taxpayers absorb losses. Communities inherit weakened services and lost trust. Power has an ugly accounting trick. It keeps the rewards close and sends the bill to people who never approved the transaction.
Justice, when it comes, comes late and diluted. A decade can pass between the first warning and the first verdict, and by then the money is spent and the public has moved on. That is why these notes are about prevention. The courts are the backstop. The early signal is the front line, and the early signal is a skill anyone can learn.
The shelf also holds the other half of the story. Two of the twelve are about builders: a small Pentagon office that changed how an institution buys, and a handful of founders who questioned the most expensive assumption in their industry and opened a new one. Prevention is only half of leadership. The other half is the nerve to build something better and the discipline to let reality test it.
The subjects of these books are not morally interchangeable. A law enforcement sting is not fraud. The stakes of military targeting dwarf those of a corporate scoring tool. The comparisons that follow are structural, not moral. My working rule is simple. Ask for the artifact before accepting the attestation, then ask two questions an audit can miss: who benefited, and who paid? That is the argument of Evidence Over Attestation, and this shelf is the reading behind it.
The four I would hand a board first
All twelve earned their place. If a board gave me time for four, these are the ones.
- When the Heavens Went on Sale, by Ashlee Vance. The best book on the shelf about innovation itself: questioning the assumption everyone treats as law, then letting a real test grade the answer.
- Dark Wire, by Joseph Cox. The clearest lesson in how trust spreads through endorsement rather than verification, and what it costs when the endpoint belongs to someone else.
- Unit X, by Raj M. Shah and Christopher Kirchhoff. Proof that an institution can change, and that the rules for buying decide what an institution is allowed to build.
- Regime Change, by Maggie Haberman and Jonathan Swan. What happens to the record, and to accountability, when authority concentrates and procedure is treated as friction.
The reading list at a glance
| Title | Author | The question it answers |
|---|---|---|
| Dark Wire (2024) | Joseph Cox | Who controls the endpoint you trust |
| Project Maven (2026) | Katrina Manson | What a human checkpoint means at machine speed |
| The Mastermind (2019) | Evan Ratliff | Who sees the whole system, and what happens if they go |
| Billion Dollar Whale (2018) | Tom Wright and Bradley Hope | How prestige stands in for verification |
| Challenger (2024) | Adam Higginbotham | How a known warning becomes normal |
| Chokepoints (2025) | Edward Fishman | Which dependency becomes leverage in someone else’s hands |
| The Fund (2023) | Rob Copeland | Whose judgment is hidden inside a score |
| Unit X (2024) | Raj M. Shah and Christopher Kirchhoff | How the rules for buying decide what gets built |
| When the Heavens Went on Sale (2023) | Ashlee Vance | Which expensive assumption is only a habit |
| Nuclear War: A Scenario (2024) | Annie Jacobsen | What must be decided before the clock starts |
| Biological War: A Scenario (2026) | Annie Jacobsen | Who acts when the signal is faint and shared |
| Regime Change (2026) | Maggie Haberman and Jonathan Swan | What happens to the record when authority concentrates |
Twelve books that changed the questions
For each book I have kept the author’s account separate from the lesson I draw from it, especially where the account rests on contested interviews.
Dark Wire, by Joseph Cox
The most effective back door was the front door of trust.
The surprise in Dark Wire is not that law enforcement read encrypted criminal messages. It is that criminals helped distribute the system that betrayed them. The FBI and the Australian Federal Police quietly ran ANOM, the phone network at the center of the FBI operation known as Trojan Shield, an encrypted phone brand that spread through trusted underworld resellers. It had everything buyers expected: handsets, updates, support and an air of security. But the operator controlled the system, and by the June 2021 takedown authorities had collected more than 27 million messages.
The technology did not need to beat cryptography. It needed users to believe the people controlling the endpoint were on their side. Buyers checked the social standing of the seller far more carefully than the technical power of the operator. That is an uncomfortable mirror for enterprise procurement, where we ask who else uses a service and which analysts endorse it, and rarely ask who can change the code underneath us, who holds the keys, and what the operator can see.
Once the sting worked, it also produced more urgent threats than the agencies could act on. Collecting evidence creates a duty to respond, and AI monitoring is arriving at the same problem.
ANOM was a lawful operation and its deception is not fraud for private gain. What it shows is how dangerous social trust becomes when an insider’s endorsement is mistaken for independent assurance. Trust is a decision to accept exposure. It is not evidence that the exposure has gone away.
The question it leaves me: which product in my stack would still be trustworthy if its operator became untrustworthy tomorrow?
Project Maven, by Katrina Manson
A human checkpoint is real only if the human can still change the outcome.
Maven began with a practical problem. Analysts were receiving far more drone video than they could watch, and computer vision promised to sort it. Manson follows the program from its 2017 start through its internal battles, its field deployment and its growth into a broad system of AI-enabled analysis, shaped by a driven Marine colonel, imperfect data, operational pressure and public controversy.
The consequential detail is not that AI made an analyst faster. It is that speed changes the task. When a model turns hours of watching into minutes of recommendations, the human at the end appears to keep authority. But a person facing an accelerating queue, with limited context and a presumption that the machine has already screened the cases, is in a different position from a person with time to think. A target detected is not a target lawfully identified. The useful measure is not how many approvals appear in a log. It is whether humans can catch and stop the mistakes the automation actually makes.
Running a system in shadow mode, alongside operations rather than inside them, is how it learns the limits of its competence before other people pay for the lesson. And the Google employee revolt belongs in the architecture story: the intended use of a technology decides whether the people able to build it will stay.
Automation can accelerate an answer without accelerating anyone’s ability to judge it. When those two speeds diverge, a human in the loop becomes a human in the paperwork.
The question it leaves me: if an agent acts a hundred times faster tomorrow, which permissions must be narrowed today so that a meaningful refusal remains possible?
The Mastermind, by Evan Ratliff
The one person who sees the whole system is both its sovereign and its weakness.
Paul Le Roux understood what good software architects understand. A distributed organization becomes enormously productive when specialized people perform narrow tasks under a common orchestrator. Ratliff follows a gifted programmer whose encrypted communications and remote operations knit together a sprawling criminal enterprise. Participants understood only their own assignment. Le Roux understood the machine.
That asymmetry gave him reach, secrecy and leverage, and it created a severe fragility. The whole system depended on one person’s private map. When law enforcement captured him and secured his cooperation, the concentration of knowledge that had protected the enterprise brought it down. This architecture organized wrongdoing, and its efficiency multiplied the harm to victims while shielding the people giving orders. Sophisticated coordination without conscience is a force multiplier for harm.
The lesson is that architecture allocates power long before any governance chart catches up. A system can look decentralized at the edges and be absolute at the center. Agent-driven organizations will meet a gentler version of this when one founder or engineer is the only person who understands the prompt chains, the credentials and the overrides. The answer is not to give everyone every secret. It is to make authority legible: durable agent identities, explicit permissions, action records, and a tested way to recover without asking one person to narrate what happened. If we cannot say who controls the root of a system, we cannot honestly say who governs it.
The question it leaves me: could a capable successor reconstruct our most important automated workflows tomorrow without asking their creator to explain them?
Billion Dollar Whale, by Tom Wright and Bradley Hope
The most prestigious name in the room is not a check. It is a reason nobody else checked.
Jho Low was in his late twenties, held no formal title at Malaysia’s sovereign wealth fund, 1MDB, and still directed the movement of billions of dollars out of it. Wright and Hope, the Wall Street Journal reporters who broke much of the story, follow the money from a 2009 joint venture through bond deals arranged by Goldman Sachs, into offshore shell companies, and out into superyachts, private jets, art, Manhattan real estate and the financing of a Hollywood film. Roughly $4.5 billion went missing.
What makes the book so instructive is that Low rarely had to defeat a control. He walked through the front door. Bond underwriters, law firms, auditors and private banks each lent their names to the transactions, and each treated the presence of the others as proof that someone had verified the underlying purpose. Low’s real skill was assembling an attestation cascade: every respected signature lowered the questions the next signer asked. The scheme unraveled because a former employee of a partner firm leaked internal emails to journalists, and reporters kept following the documents when institutions stopped asking.
The legal tail is long. Goldman Sachs paid a record settlement, two of its bankers were convicted, and in December 2025 Malaysia’s High Court sentenced former Prime Minister Najib Razak to a further fifteen years. Low remains a fugitive. Justice arrived, but a decade after the first warning, and diluted by the time it landed.
More validators can mean less validation when each one assumes the others have already checked.
The question it leaves me: in my own chain of approvals, which signature is standing in for a check nobody actually performed?
Challenger, by Adam Higginbotham
A warning that has been true too many times without catastrophe stops sounding like a warning.
Challenger broke apart seventy-three seconds after launch on January 28, 1986, and seven crew members died, including the teacher Christa McAuliffe. Higginbotham’s history is the fullest account yet of how it happened, and it is less a story of one bad night than of years of reasonable-sounding decisions. Erosion of the rubber O-rings in the solid rocket booster joints had been seen on flights as early as 1981. Each time the shuttle came home safely, the anomaly was reclassified a little further toward acceptable.
The night before the launch, engineers at the booster contractor, Morton Thiokol, warned that the seals had never been tested at temperatures that cold and recommended against flying. Under schedule pressure, the burden of proof quietly flipped. Instead of being asked to show it was safe to launch, the engineers were asked to prove it would fail, which they could not do with the data they had. Their own managers were asked to set aside their engineering judgment and make a management decision. The launch went ahead in freezing weather.
The Rogers Commission that followed, and the physicist Richard Feynman’s simple demonstration with a piece of O-ring and a glass of ice water, made the cause plain. The sociologist Diane Vaughan later named the pattern the normalization of deviance. Higginbotham honors the engineers who raised their hands and shows how carefully built institutions can still talk themselves out of hearing them.
The most dangerous moment in a review is when the burden of proof moves from the people who want to proceed to the people who are worried.
The question it leaves me: which known anomaly in my own systems have we stopped discussing because it has never yet caused a failure?
Chokepoints, by Edward Fishman
A dependency is a strategic choice even when nobody remembers making it.
Fishman, who helped design sanctions at the State Department and the Pentagon, shows how global integration created narrow passages through which extraordinary power can be exerted: dollar finance, payment messaging, advanced semiconductor equipment. The power comes less from size than from the inability of others to route around a node quickly.
The book’s value for technology leaders is its attention to the second move. The first use of a chokepoint can be devastating. It also teaches the target to build an exit, so leverage has a half-life whenever the people who feel it can invest in escaping it. The commercial parallel is immediate. A company can claim a multi-vendor strategy while every vendor depends on the same identity provider, chip supply or model API. Diversity on the invoice can hide concentration two layers down.
A chokepoint is a dependency you discover at the moment someone else changes the terms. Good governance discovers it earlier, and a contingency plan is a tested failover, not a second logo on a slide. This book is also the foundation for The Chokepoint Thesis, the research I built after reading it.
The question it leaves me: which supposedly redundant service in our business would fail with the same upstream provider, and when did we last rehearse the exit?
The Fund, by Rob Copeland
Every score has an author, even when the interface hides the signature.
Ray Dalio built an extraordinary investment firm and promoted a philosophy of radical truth and written principles. Copeland examines what those ideals could feel like inside Bridgewater, where meetings were recorded, colleagues rated one another, and quantified traits shaped judgments about credibility. His account is sharply critical. Dalio and Bridgewater have disputed substantial parts of it, and their response deserves to be read alongside the book.
The governance question does not require settling every contested episode. Picture any organization that scores colleagues, ranks whose judgment counts, and treats the resulting dashboard as more objective than ordinary disagreement. Who chose the weights? Who hears an appeal? Can the system produce a conclusion that the person who commissioned it must accept? Once judgment acquires numbers, people stop seeing the judgment inside it. A score of 94 feels more scientific than a written disagreement, even when the formula is a compressed record of preferences and power.
I think about this every time I review an AI evaluation harness or an executive dashboard. A leader can build a machine that agrees with him at scale and mistake consistency for independence. If a metric cannot embarrass its sponsor, it is probably measuring obedience as much as performance. That line is my inference, not a quotation from either side, but it is the test I take from their dispute.
The question it leaves me: have we written down, in advance, what number on our dashboard would mean our strategy is failing, and is there someone besides the strategy’s sponsor who is allowed to act on it?
Unit X, by Raj M. Shah and Christopher Kirchhoff
The rules for buying technology quietly decide which future an institution is allowed to build.
Two insiders describe the effort to connect the Pentagon’s procurement machinery with the faster cycles of commercial technology. The Defense Innovation Unit did not begin by inventing a new drone. It began by changing the path through which a government customer could state a problem, find a supplier, sign an agreement and field something useful.
That sounds less glamorous than the products. It is often more consequential. A procurement process built around multiyear specifications will exclude smaller companies that iterate software in months, so the procedure has already chosen which firms can participate. Process becomes policy by deciding the possible outcomes before anyone evaluates a prototype. Shah and Kirchhoff write from inside the reform, so the book carries both valuable detail and the view of people defending a project they built. Read beside Project Maven, it shows that innovation needs urgency and champions, and also real users, repeatable authority, budgets and evidence that the fielded system works.
This book is a counterweight to the darker shelf. Not every use of influence is corrupt. People can use access and determination to remove waste and serve a public purpose. One caution belongs here. Faster acquisition is a clear good when the capability is defensive. The same speed lowers the threshold for capabilities whose use is far more debatable, and a reform that accelerates the process has to ask what the process is accelerating. The bureaucracy you cut determines the errors you stop catching.
The question it leaves me: are we helping a customer clear a real institutional obstacle, or demonstrating technology on the wrong side of it?
When the Heavens Went on Sale, by Ashlee Vance
The fastest way to change an industry is to ask why its most expensive assumption is true.
This was my favorite book of the twelve. Vance, a Bloomberg journalist, spent years with four unconventional companies trying to make space cheap. Planet Labs began when a few NASA engineers asked why a satellite had to cost hundreds of millions of dollars, then built small ones from consumer-grade parts and launched enough of them to photograph the Earth’s land every day. Rocket Lab grew out of Peter Beck, a self-taught engineer in New Zealand without a university degree, who built a small rocket with 3D-printed engines and reached orbit in 2018. Astra raced to mass-produce rockets and learned hard lessons in public. Firefly went bankrupt, was revived by a Ukrainian investor, and was then reshaped again under government scrutiny.
The book is about innovation in its most honest form. Every one of these founders started by questioning an assumption the incumbents treated as physics: that satellites must be large, that launch must be rare, that space belongs to governments and primes. Some of those assumptions turned out to be habits rather than laws, and the companies that saw it first opened an industry. Vance is equally clear that questioning assumptions is only half the job. The other half is the test. A rocket reaches orbit or it does not, and no investor deck changes the result.
What separates the survivors is how they treated failure. The teams that instrumented everything, found the root cause and fixed it moved forward. The teams that moved fast without learning from each attempt repeated their failures on a faster schedule. Speed is a gift when every failure changes the design, and a cost when it does not. It is the best description I have read of what a founder working with AI now faces: cheap iteration, enormous possibility, and a physical world that still grades the work.
Question the expensive assumption, then let reality grade the answer.
The question it leaves me: which assumption in my market is everyone treating as a law, and what is the real-world test that would prove it is only a habit?
Nuclear War: A Scenario, by Annie Jacobsen
Every safeguard that matters in a crisis was decided before the crisis began.
Jacobsen builds a minute-by-minute scenario from declassified documents and interviews with former officials, commanders and scientists. A single surprise launch reaches American early warning satellites, and from that moment the reader watches a chain of decisions made on a clock measured in minutes. The president has only a few of them to evaluate ambiguous data and choose among options prepared years earlier. A response is misread by another nuclear power, and the escalation runs on. The whole arc takes about seventy-two minutes.
Some experts dispute the assumptions that keep her escalation uninterrupted, and I read it as a stress test, not a forecast. As a stress test it is unforgettable. It shows what happens when a decision cycle runs faster than the human cycle of judgment. Nobody in the scenario has time to invent a policy, verify a sensor or debate a principle. Everything that mattered was written down, rehearsed and committed to long before the alarm.
The parallel to automated systems is direct and sobering. In cyber defense, trading and agent workflows, some actions become effectively irreversible the moment they fire. The time to design the brakes, the human checkpoints and the conditions for stopping is before the system is live, not during the incident.
A safeguard invented during the emergency is a hope, not a safeguard.
The question it leaves me: which of our automated actions cannot be undone once triggered, and when did we last rehearse stopping them?
Biological War: A Scenario, by Annie Jacobsen
A threat that spreads quietly defeats institutions that wait for certainty.
Jacobsen applies her scenario method to an engineered plague that escapes from a laboratory and travels with ordinary people across airports, cities and borders. The early cases look like something familiar, and the first responders reasonably treat them that way. Agencies hold fragments of the picture and do not share them quickly. Specialized response units need days to mobilize. By the time the pattern is clear, the problem has moved past the point where any one institution could contain it.
The contrast with Nuclear War is the reason to read them together. A missile launch is unmistakable and belongs to a single command authority. A spreading threat arrives as noise, across many institutions, each of which can rationally wait for someone else to declare the emergency. Jacobsen’s real subject is the gaps between systems. No single failure causes the catastrophe. The handoffs do. Reviewers differ on how far the scenario pushes its worst case, and it is best read, like its companion, as a stress test of how institutions coordinate.
The lesson travels well beyond biology. A slow data leak, a model that drifts quietly, or a supply chain compromise spreads the same way: through scattered signals that each look minor, owned by teams that each see only part of the pattern.
Waiting for certainty is itself a decision, even when nobody signs it.
The question it leaves me: when a problem shows up as faint signals across several teams, who has the explicit authority to act before the picture is complete?
Regime Change, by Maggie Haberman and Jonathan Swan
An exception without an expiry date becomes a policy without an author.
Haberman and Swan report on the first year of Donald Trump’s second presidency and the people, loyalties and structures around it. The book rests on extensive access reporting and its interpretation is politically disputed. Readers should weigh the accounts, the incentives of the participants and the responses of those criticized. The constitutional stakes are real and specific, not a management parable. What interests me for this project is one pattern inside those stakes: what happens when decision authority concentrates and procedural friction is treated as an obstacle to sweep aside.
Formal processes are carried out by particular people who decide whether to ask the second question, record a dissent or refuse a requested act. When decisions move from documented channels into private conversations, the institution gains speed and loses the ability to establish later what was authorized and why. Institutions need room to act in emergencies. They also need to know when the emergency ends. A permanent exception is a new rule without the honesty of an open rulemaking, and the same pattern appears in technical governance when a team widens an agent’s privileges temporarily or waives a deployment test because a deadline is near.
The decisive measure of accountability, in this book as in The Fund and The Mastermind, is the fate of an inconvenient objection. A system that allows dissent only when the powerful agree with it has not distributed judgment. It has staged consultation.
The question it leaves me: when authority overrides a safeguard, can a later reviewer find the request, the reason, the objections, the decision and its expiration date?
The arguments these books have with one another
Reading twelve books as twelve summaries would miss the best part. They disagree across subject, era and moral position, and the friction teaches more than any list of principles.
The profit in dishonesty. If deception pays reliably and candor is punished, dishonesty stops being an individual defect and becomes a behavior the institution is training people to repeat. Billion Dollar Whale shows how far that can run before anyone stops it. The surest evidence of institutional character is what gets rewarded when the transaction is profitable and the truth is costly.
The cost that never reaches the ledger. The gain is visible and the loss is dispersed. A taxpayer never sees the chain from a favored deal to a weakened service. The employee who objected lost the opportunity long before the scandal broke. The wrongdoer gets a transaction. Everyone else gets a consequence.
Borrowed certainty. ANOM traveled through trusted distributors. 1MDB moved through prestigious banks and auditors. Bridgewater’s disputed scoring raises the question of whether numbers can lend authority beyond what the observations deserve. More validators can mean less validation when each one assumes the others have already checked. Somebody must own the artifact rather than the handoff.
The missing time. Maven asks what oversight means at machine speed. Nuclear War shows a clock measured in minutes. Biological War shows the opposite problem, a signal so slow and scattered that nobody feels entitled to act. Unit X asks how to remove delays that add nothing. The distinction is between dead time, waiting caused by an avoidable handoff or a queue nobody owns, and thinking time, when an expert notices an assumption or declines a harmful action. An efficient institution knows which is which.
The test that cannot be argued with. The Fund shows a score with no path to correction. When the Heavens Went on Sale shows the opposite: a rocket reaches orbit or it does not, and the founders who learned from every failure moved ahead. Every product needs its equivalent of a launch, a real test that can disappoint the people who built it.
The warning that became wallpaper. Challenger gives this pattern its sharpest form. An anomaly seen on flight after flight became normal because the shuttle kept coming home. The same thing happens in softer form everywhere: small model errors accepted because the product still ships, repeated overrides accepted because the urgent project succeeded. Familiarity gradually impersonates evidence.
The hidden sovereign. The Mastermind shows the operator with complete knowledge. Chokepoints shows the gatekeeper everyone must pass. Regime Change shows procedure under concentrated authority. Having visibility, holding technical control and holding legitimate authority are three different things, and the most powerful party in an organization may be whoever holds the credentials, sets the weights or approves the exceptions.
The early signals
Prevention is a skill, and it comes down to recognizing a pattern while it is still small. Each of these shows up in an ordinary week, long before anyone would call it a scandal. Here is what each looks like at that stage, the question that exposes it, and the book where it grew to scale.
- A name is doing the work of a check. It sounds like “they use it too” or “the big firm signed off.” Ask: who looked at the artifact itself, and what did they see? Dark Wire and Billion Dollar Whale.
- One person or service can see the whole system and nobody else can. It sounds like “ask him, he set it up.” Ask: if that person were gone tomorrow, who could reconstruct this? The Mastermind.
- A score has never once embarrassed the person who commissioned it. It sounds like “the dashboard has been green all quarter.” Ask: what number would have to appear here for us to change course, and has anyone written that number down? The Fund.
- An exception has no expiry date and no author. It sounds like “we turned that check off during the launch.” Ask: who approved it, what was the objection, and when does it end? Regime Change.
- A known anomaly has stopped being discussed. It sounds like “that always happens, it has never caused a problem.” Ask: what would we need to see to prove it is safe, rather than assuming it is? Challenger.
- Speed is being measured, but nobody can say which minutes were thinking time. It sounds like “we cut review from a week to a day.” Ask: which delay added knowledge, and which was waste? Project Maven and Unit X.
- A supposedly redundant service shares an upstream. It sounds like “we have two vendors for that.” Ask: do both depend on the same gate two layers down, and when did we last rehearse the exit? Chokepoints.
- A problem is visible only in pieces, and nobody owns the whole. It sounds like “that is probably someone else’s issue.” Ask: who has the authority to act before the picture is complete? Biological War.
None of these proves anything on its own. Each is a reason to ask one more question while the question is still cheap. That is the whole skill.
Take the eight signals with you: a one-page worksheet (PDF).
Download the worksheet (PDF)The evidence test I now take into a meeting
I would rather leave a meeting with a handful of hard artifacts than a binder of confident assurances. The NIST AI Risk Management Framework’s four functions, govern, map, measure and manage, give the structure. The table makes it concrete for systems that can act.
| What I am being told | What I ask to see | Book anchor |
|---|---|---|
| Our vendor and data are secure | Key custody, endpoint access, update authority, and independent testing records | Dark Wire; Billion Dollar Whale |
| The agent follows policy | Versioned permissions, explicit prohibitions, action and spending limits, and a named accountable owner | Project Maven; Nuclear War |
| A human approves important actions | Queue volume, time and context per review, real rejection rates, and evidence that reviewers can stop execution | Project Maven; Challenger |
| The workflow is under control | Durable agent identities, separation of duties, an action log, and a tested recovery path | The Mastermind; Biological War |
| We have multiple providers | A dependency graph and a demonstrated exit or failover, including upstream services | Chokepoints |
| The model scored well | The dataset, the weighting choices, known failures, independent challenge cases, and a threshold that can block release | The Fund |
| It worked in our demonstration | Tests under real production conditions, adversarial cases, drift evidence, incidents, and observed cost | When the Heavens Went on Sale; Unit X |
| Leadership can authorize exceptions | A dated override record stating who acted, why, what objections existed, and when the exception expires | Regime Change; Challenger |
None of these artifacts proves a system permanently safe. They make its claims contestable and its behavior reconstructible. Documentation alone does not produce integrity, since an institution can keep perfect records of misconduct and still reward the person responsible, so I add three questions to every review. Who stands to profit if we accept this assurance? Who absorbs the harm if it is false? Is the person with the most to gain also controlling the evidence used to approve the decision?
One principle joins all eight tests. The evidence must keep the ability to disappoint the person who asked for it. Without that, measurement only reassures.
The lesson I carry forward
A good dashboard lets people see what would otherwise be lost in noise. I am not arguing against that. I am arguing for the demanding relationship it must keep with reality.
At the end of a long meeting, a leader wants a coherent story. The most valuable thing anyone can bring into that room is the piece of evidence that makes the story less coherent: a failed field test, an unexamined root credential, a payment with no purpose, a scorched seal from the last flight. Such evidence may embarrass the sponsor and delay the launch. That is an expensive kind of honesty. It is also the only kind that compounds.
An organization reveals its real culture at the moment the uncomfortable artifact arrives. Do its leaders ask to see it? Does the person who found it get a fair hearing? Does the record survive the meeting? Does the next decision change? A printed value statement cannot answer those questions. The room will.
Twelve books have strengthened my belief that a trustworthy system needs a written boundary around authority, testing against real conditions, a record that survives its creators, and a way to correct the people in charge. They also left me with the reminder from the best of them: the builders who changed their industries did it by questioning what everyone else assumed and then letting reality grade the work. Prevention and invention are the same discipline pointed in two directions.
The next twelve are already on the shelf. The Next Leadership Dozen moves from how systems fail to the decisions made before any evidence exists: who decided what success means, what gets measured, and which consequences count.
The ten lessons, once more
Trust is a decision to accept exposure. Whoever controls the root controls the system. Speed moves judgment upstream. A score is only as honest as the person who set the weights. A dependency is a strategic choice even when nobody remembers making it. An exception without an expiry date becomes a policy without an author. A warning that has been true too many times stops sounding like a warning. The rewards concentrate and the costs disperse. Question the expensive assumption, then let reality grade the answer. Character is what you do with power when the easier path pays.
The question I carry into every board discussion and every AI deployment is this one. If the system were wrong in a consequential way, how would we know, who could say so, who would profit from silence, and who would pay the price? The answer is not on the dashboard. It is in the evidence the institution is willing to let speak. If you have read any of the twelve, I would like to know which one you would hand a board first.
Books and verification notes
The lines in bold are my own interpretations unless attributed. The book sections describe my readings of the works, not an adjudication of every contested account.
- Cox, Joseph. Dark Wire: The Incredible True Story of the Largest Sting Operation Ever. PublicAffairs, 2024. Operation Trojan Shield facts checked against the US Department of Justice announcement of June 8, 2021.
- Manson, Katrina. Project Maven: A Marine Colonel, His Team, and the Dawn of AI Warfare. W. W. Norton, 2026. NPR interview with the author, March 2026.
- Ratliff, Evan. The Mastermind. Random House, 2019.
- Wright, Tom, and Bradley Hope. Billion Dollar Whale. Hachette Books, 2018. Najib Razak’s December 2025 conviction and sentence: TBS News.
- Higginbotham, Adam. Challenger: A True Story of Heroism and Disaster on the Edge of Space. Avid Reader Press, 2024. Report of the Presidential Commission on the Space Shuttle Challenger Accident, 1986; Diane Vaughan, The Challenger Launch Decision, 1996.
- Fishman, Edward. Chokepoints: American Power in the Age of Economic Warfare. Portfolio, 2025.
- Copeland, Rob. The Fund: Ray Dalio, Bridgewater Associates, and the Unraveling of a Wall Street Legend. St. Martin’s Press, 2023, read alongside Bridgewater’s November 2023 public response.
- Shah, Raj M., and Christopher Kirchhoff. Unit X: How the Pentagon and Silicon Valley Are Transforming the Future of War. Scribner, 2024.
- Vance, Ashlee. When the Heavens Went on Sale: The Misfits and Geniuses Racing to Put Space Within Reach. Ecco, 2023.
- Jacobsen, Annie. Nuclear War: A Scenario. Dutton, 2024.
- Jacobsen, Annie. Biological War: A Scenario. Dutton, 2026.
- Haberman, Maggie, and Jonathan Swan. Regime Change: Inside the Imperial Presidency of Donald Trump. Simon & Schuster, 2026.
- National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0). US Department of Commerce, 2023.
Related: The Leadership Dozen: Twelve Books for the Intelligence Era, The Next Leadership Dozen, Evidence Over Attestation, Part One and the twenty-one-point standard.