Information Gradients: Why the Truth Gets Thinner as It Climbs

The board hears a color

Late in the night shift, a board operator silences an alarm without looking at it, because that alarm has been crying wolf for a year and the real ones are somewhere in the same flood. He knows which interlock has been jumpered since the last startup and never came back in. He knows the variable that drifts past its limit on hot afternoons and walks itself back before anyone logs it. None of this is a secret to him. It is the texture of his job. Three months later the unit lets go, an investigation convenes, and the report notes with some surprise that the warning signs were present. They were always present. They were sitting in the head of a man who had no safe, rewarded, or even available way to say them out loud to anyone who could act.

We have watched the equivalent of that scene in a hospital, a trucking dispatch office, a bank's underwriting floor, a portfolio company, and the third quarter of a billion dollar transformation. The vocabulary changes. The shape does not. The single fact that would have predicted the failure existed, on time, in a specific person's firsthand experience, and the organization's instruments did not contain it. This is not bad luck and it is not incompetence. It is structural, and it runs the same way in every industry. We call it the information gradient: ground truth is richest at the floor and gets thinner, safer, and more flattering at every level it climbs, until what reaches the top is a color on a slide.

What the floor knows and the system does not

Start with the most studied version, because safety has spent half a century mapping it. James Reason's organizational accident model^[1] separates the unsafe act at the sharp end from the latent conditions upstream that made it likely: the staffing call, the deferred maintenance, the goal conflict written into a budget. The act is visible. The latent condition is not, because it lives in the experience of the people who absorbed it daily until they stopped noticing. Sidney Dekker calls this the gap between work as imagined and work as done^[2], and his central point is that the gap is permanent. Procedures describe the work management believes is happening. The work actually happening, with all its quiet substitutions, lives only in the heads of the people doing it. Ask them to walk through a normal messy task and the gap appears in the first thirty seconds. No system records it, because the system was built to hold work as imagined.

The most disturbing mechanism in the canon is Diane Vaughan's normalization of deviance^[3], from her reconstruction of the Challenger decision. Each flight that came back with O-ring damage made the next anomaly a little more acceptable. The baseline moved. What would have stopped the launch one year became routine the next, through a sequence of individually rational choices. A drifted baseline is invisible from inside; nobody experiences it as deviance anymore. The only way to find it is to ask a long tenured person what would have stopped the job a decade ago that does not stop it today, and then listen to what they laugh off. The laugh is the data, and it is in no record, because to the person living it there is nothing to record.

This is why driving down the easy metric does not buy down the catastrophic risk. The Krause Bell Group's work on serious injuries and fatalities^[4] found that fatal events have different precursors than minor ones, so a site can push its recordable injury rate to an enviable low while leaving its fatality exposure untouched. Texas City had excellent personal-injury numbers and a catastrophic process-safety culture. The precursor of the next death is a task with high energy present and a control that has quietly degraded, and the only people who can see that combination before it kills someone are the ones standing next to it. They can see it for months; the metric leadership watches cannot see it at all. The organization measures what is easy to roll up, and the thing that predicts the disaster does not roll up.

Why the truth dies on the way up

Now the engine underneath every example. The information does not fail to reach the top because the channel is noisy. It fails because every link in the channel has a private reason to soften the signal. The frontline worker who reports a near miss invites paperwork, a drug test, or blame, so the story stays in the break room. The supervisor who carries an expensive problem upward learns from one bad meeting what is safe to bring next time. The function packages its workstream to look governed, because looking governed is its product. By the time the signal reaches the dashboard it is a category, and by the board it is a color. No one in that chain lied. Each did the sensible thing for their position, and the sum of sensible local moves is a decision maker who is structurally the last to know the one thing the floor knew first.

Transformation programs gave this a name. Practitioners call it the watermelon problem: status reports that are green on the outside and red on the inside. Mark Keil and colleagues studied status misreporting directly^[5] and found what they call the mum effect: across study after study, people reliably withhold bad news about a project, and auditors whose literal job is to report it often do not either. The reason is the same everywhere: reporting red marks you as the problem. Their fix is not exhortation; it is an independent reporting line and making red a safe color, which is to say, removing the punishment that creates the gradient. Until you do that, every layer of review adds optimism rather than accuracy, and the program that is quietly failing keeps passing its own gates.

The gradient is not only about bad news traveling poorly. Sometimes the basic message never arrives at all. Donald Sull's decade-long study of strategy execution^[6], built on a survey of 7,600 managers across 262 companies, asked people simply to list their company's top priorities in their own words. Only 55 percent of middle managers could name even one of the top five, and that share collapses further down the line. In one firm whose engagement survey reported that 84 percent of staff were clear on priorities, fewer than a third of the top team could name two of the five. The survey said the message had landed; the behavioral test said it had not. This is the gradient on a different axis: the company believed it was aligned because it measured the espoused theory, the answer to a leading question, rather than what people actually do on Tuesday.

The same shape with money on it

Private equity is where the gradient gets priced, because a buyer is wiring money against a story. The confidential information memorandum is, in the trade's own phrasing, an advocacy document: the forecast stretched to sell, the wins all credited to current management, the close that looks like a process rather than one controller's heroic spreadsheet. The equity case actually rests on facts two and three levels down, where the CIM has no access and no incentive to look. Geoff Smart's firm built its assessment method^[7] on exactly this asymmetry, and found that unstructured judgment of a management team succeeds about as often as a coin flip, around 50 percent, while a structured chronological interview that forces specific recent episodes and checks them against named references gets to roughly 90 percent. The difference is entirely in refusing the polished summary and demanding the moment behind it. The skip-level manager knows which executive is carried by a strong deputy. The org chart records the title. Only one of those facts survives the deal.

Artificial intelligence is the most expensive contemporary proof of the gradient, and it inverts the usual moral. MIT's Project NANDA found that about 95 percent of enterprise generative-AI pilots show no measurable return^[8] on some thirty to forty billion dollars of spend. A leading cause is that programs are scoped off job descriptions and management's picture of the work, not the real task composition that only the people doing the work can supply. Meanwhile the value is hiding underground. A KPMG study of 48,000 people across 47 countries found that 57 percent hide their AI use^[9] and present the output as their own. Ethan Mollick named these people secret cyborgs^[10] and is precise about why they stay hidden: bans push the work onto personal devices, AI-touched output gets judged worse once people know, and disclosing that you automated most of a task invites a cut. So the organization spends tens of billions rebuilding, badly, capabilities its own staff have already proven and concealed. The proof of which AI bets pay off sits in the same heads as everything else, kept there by the same punishment structure, and you cannot survey it out of people because the survey is what they are hiding from.

Read across these worlds and the convergence is the whole argument. McKinsey finds even self-declared successful transformations capture only about 67 percent of their potential value^[11], with the rest lost first at target-setting, on a number nobody on the floor believed, then during an implementation the floor could see stalling. The numbers differ by sector and the mechanism is identical, which is what you expect from something structural rather than circumstantial. A safety record, an engagement score, a status color, a pipeline coverage ratio, a pilot's projected savings: each is an honest measurement of the wrong thing, produced by a chain that thins the signal as it climbs.

Reading the gradient instead of the rollup

Notice what every expert above is prescribing, underneath the framework names. Reason traces the act down to the latent condition. Dekker walks the messy task with the person who does it. Vaughan asks the long tenured worker what used to stop the job. Keil installs an independent reporting line and makes red safe. Smart demands the specific recent episode and checks it against a named reference. Mollick removes the penalty and watches what surfaces. These are not six methods but one move in six costumes: get a neutral party to ask a specific person about a specific recent moment, somewhere the honest answer cannot be used against them. The reason organizations do not have the answer is almost never that it is unknowable. It is that the only channels built for asking are the ones the floor learned long ago to lie into.

The optimistic reading, and we hold it firmly, is that the gradient is a property of the instruments, not the people. The frontline operator is not withholding; he is responding rationally to a channel that has punished candor. Change the channel and the information is right there, the same information that was always there, available in time to act on rather than in time to explain. A survey locates the sick unit and stops. An exit form collects the safe answer by design. A dashboard rolls up what is easy to roll up. None of them was built to climb back down the gradient and retrieve the fact at the altitude where someone actually knows it.

That is the instrument the evidence has described for fifty years and few have operationalized at scale: a neutral, confidential conversation, anchored to a real recent moment rather than a feeling on demand, run for enough people that no answer traces back to one person, and read at the altitude where each person genuinely knows something. It is the kind of thing we are building at Latent Variables. The literature already told us where the truth sits. The open problem was only ever reaching it before the slide turns green and the unit lets go.