The Confidence Machine

By John Carmean

May 19, 2026

A recurring forensic profile runs through the major intelligence failures of our last century. The signals were public. The analysts were employed. The integration was missing.

The pattern is consistent enough that it has its own literature, its own commissions, its own postmortems running back nearly a century. The names of the events change. The conclusion does not. After every catastrophic surprise, an inquiry convenes, examines the record, and produces a phrase that has become almost a cliché in the discipline: the warnings were there.

They usually were. What was missing was a method for surfacing the disagreement between them. That method is the actual intelligence. Everything else is reporting.

The Pearl Harbor study that named the problem

The foundational text in the field is Roberta Wohlstetter's Pearl Harbor: Warning and Decision [1], published in 1962 and awarded the Bancroft Prize the following year. Wohlstetter was a RAND historian who spent several years reading the thirty-nine-volume record of the Joint Congressional Committee on the Investigation of the Pearl Harbor Attack [2].

Wohlstetter's thesis was unsettling. The United States did not lack information about Japanese intentions in late 1941. It had broken the Japanese diplomatic codes. It had intercepts. It had naval movement reports. It had the negotiations breakdown over the oil embargo.

What it lacked was the ability to separate signal from noise. Her conclusion, in the line that has been quoted in intelligence training ever since, was this:

“We failed to anticipate Pearl Harbor not for want of the relevant materials, but because of a plethora of irrelevant ones." [1]

The warnings were buried inside competing and contradictory signals, none of which had been integrated against any of the others.

Wohlstetter's framework, signal versus noise, has become the standard vocabulary for intelligence postmortems. But its more important contribution was the structural claim: that the failure mode was not collection. It was integration.

Yom Kippur, 1973, and the "Concept"

More than three decades later, Israel produced its own version of the same lesson, and gave it a different name.

In the weeks before October 6, 1973, Israeli military intelligence received multiple converging warnings of an imminent Egyptian-Syrian attack. King Hussein of Jordan flew secretly to Tel Aviv on September 25, 1973 to warn Prime Minister Golda Meir personally that Syrian forces were moving into forward attack positions on the Golan front. Aerial reconnaissance showed Egyptian troop and bridging-equipment concentrations on the Suez Canal that were inconsistent with the exercise the Egyptians publicly claimed to be running. A junior intelligence officer at AMAN's Southern Command, Lieutenant Benjamin Siman-Tov, filed a memo on October 1 warning that the Egyptian buildup was preparation for war, not training. He was overruled by his superior, who declared it a "low probability." The original memo was never escalated. The Yom Kippur War began five days later.

The Agranat Commission, chaired by Supreme Court President Shimon Agranat and convened on November 21, 1973, produced an interim report on April 1, 1974 that identified what had gone wrong [3]. It coined a term that has since entered the Israeli political and intelligence lexicon: ha-konseptzia, the Conception. Israeli military intelligence, the commission found, was operating under a fixed analytical framework that held Egypt could not and would not initiate a war it could not win. Data points that disagreed with the Conception were systematically reinterpreted to fit it. The data was integrated. The integration was wrong.

The most rigorous academic treatment of the failure is Uri Bar-Joseph's The Watchman Fell Asleep [4], which reconstructs the chain of overridden warnings in granular detail. Bar-Joseph's argument extends Wohlstetter's: the failure was not noise drowning out signal. It was a method that actively suppressed signals that disagreed with the consensus reading.

A sharper claim and also the correct one for our purposes.

The August 6 brief, the Phoenix memo, and a system that could not integrate

The most exhaustively documented intelligence failure of the last twenty-five years is the one that produced The 9/11 Commission Report, published in July 2004 [5]. The Commission spent two years reconstructing the warning record. The record is brutal.

On August 6, 2001, the President's Daily Brief carried an item titled "Bin Ladin Determined to Strike in US." It was declassified by the White House on April 10, 2004, and is reproduced in full as an exhibit to the Commission report. It stated that the FBI was running approximately seventy full field investigations of Bin Ladin-related activity inside the United States, and that there were "patterns of suspicious activity in this country consistent with preparations for hijackings or other types of attacks." [6] That brief was thirty-six days before the attacks.

On July 10, 2001, FBI Special Agent Kenneth Williams in the Phoenix field office had filed what is now called the Phoenix Memo, recommending that the Bureau investigate "the possibility of a coordinated effort by Usama Bin Ladin"‍ ‍to send operatives to U.S. civil aviation universities and colleges, and naming specific individuals already under Phoenix-area investigation for ties to extremist organizations [7]. The memo was not acted on. The Commission Report devotes substantial attention to it (see Chapter 8, "The System Was Blinking Red") [5].

The 9/11 Commission's conclusion, after enumerating dozens of similar pieces of information that existed in the system but were never integrated, was that the failure was not collection. It was, in the Commission's phrase, "a failure of imagination" and a failure of organization [5]. The CIA had pieces. The FBI had pieces. The FAA had pieces. The NSA had pieces. The pieces did not communicate. There was no method by which an analyst sitting in one agency could see that what they were reading agreed with, or contradicted, what an analyst in another agency was reading. The signals were present. The integration architecture was not.

Silicon Valley Bank, 2023, and supervisors who could not see across silos

The most recent canonical example, and the one with the cleanest paper trail, is the Federal Reserve's own post-mortem on the collapse of Silicon Valley Bank.

On April 28, 2023, Vice Chair for Supervision Michael S. Barr released Review of the Federal Reserve's Supervision and Regulation of Silicon Valley Bank [8]. It is a 114-page document, written under Barr's name, that constitutes the Federal Reserve's formal accounting of its own failure. The report's first key finding is unambiguous: "Silicon Valley Bank's board of directors and management failed to manage their risks." The second is more interesting: "Supervisors did not fully appreciate the extent of the vulnerabilities as Silicon Valley Bank grew in size and complexity."

The Barr Report documents that Federal Reserve examiners had identified concentration risk, interest rate risk, and management deficiencies at SVB in 2021 and 2022. They had issued formal supervisory findings. The findings were not integrated against the bank's deposit concentration, its uninsured-deposit ratio, or its held-to-maturity securities portfolio. Each examiner team saw its piece. The supervisory process did not turn the pieces into an urgent, escalated picture. The bank failed on March 10, 2023, in the largest bank collapse since Washington Mutual in 2008.

The warnings were there.

Texas, February 2021, and a 2011 report nobody read

The same pattern plays out in non-financial domains with the same regularity.

On November 16, 2021, the Federal Energy Regulatory Commission and the North American Electric Reliability Corporation jointly released the final 300-page report on The February 2021 Cold Weather Outages in Texas and the South Central United States [9]. Winter Storm Uri had caused the largest manually controlled load-shedding event in United States history. ERCOT had ordered twenty thousand megawatts of rolling blackouts. Hundreds of Texans died. The grid came within minutes of full collapse.

The most damning sentence in the report concerns a prior report, not the current one. FERC and NERC had conducted a joint inquiry after the February 2011 cold weather event in Texas. That inquiry produced detailed recommendations on generator winterization, gas-electric coordination, and cold weather preparedness. The 2021 report found, on page 171, that "Over 40 percent of the GOs/GOPs in the south central U.S. regions where 'freezing issues' were identified as the predominant cause of unplanned generation outages, derates or failures to start (ERCOT and MISO South) stated that they did not incorporate specific generator-related recommendations from the 2011 Report." [9] FERC Chairman Rich Glick said the quiet part out loud at the public presentation: "This is a wake-up call for all of us. There was a similar inquiry after Texas experienced extreme cold weather in 2011, but those recommendations were not acted on."

A complete intelligence product, sitting in the public domain for ten years, ignored. The warnings were there.

What the discordance-detectors did

The pattern is consistent enough that the affirmative case is worth stating. There have been people, in every era, who succeeded at intelligence in part because their method foregrounded disagreement.

At RAF Medmenham, the Allied photo interpretation unit during the Second World War, a WAAF officer named Constance Babington Smith ran the aircraft section. In November 1943, she was asked to look for evidence of unusual activity at a German research station on the Baltic coast called Peenemünde. The reconnaissance photographs were not yielding anything the Industry section at Medmenham recognized. Her order, she later wrote in her memoir Evidence in Camera [10], was to search for "anything queer." On a Mosquito sortie of November 28, 1943, flown by Squadron Leader John Merifield of No. 540 Squadron, she identified a small cruciform shape on a ramp that did not match anything in the existing aircraft recognition library. She insisted, against the initial dismissal of her interpretation by the Industry section, that the shape was a pilotless aircraft on a launching ramp. It was the V-1. Her identification became part of the intelligence basis for Operation Crossbow, the Allied bombing campaign against the V-weapon launch sites.

What Babington Smith did was not collect information. The photographs already existed. What she did was notice that one signal disagreed with the consensus interpretation, and refuse to reconcile it back into the consensus.

The codebreakers at Bletchley Park, whose work is documented in F. H. Hinsley's five-volume official history British Intelligence in the Second World War [11], built the same principle into their architecture. The Y Service, the network of British listening stations, was deliberately structured so that multiple stations and multiple operators heard the same intercepts and noted the same discrepancies in German operational behavior. Volume of collection mattered. Structural redundancy mattered more, because it forced signals to disagree with each other before they were trusted to agree.

That is what intelligence is, when it works.

Why this matters now

The relevant fact about most systems of professional analysis in 2026, financial, geopolitical, energy, regulatory, is that they have moved decisively in the opposite direction. The data sources that fund institutional research have consolidated. The handful of bank research desks that produce the consensus reading draw on increasingly overlapping earnings models and a narrowing pool of input series. Risk dashboards across the industry have converged toward a common architecture, with similar factor decompositions and shared scenario libraries. When the financial press aligns around a narrative, whether it is "the Fed has pivoted," "the AI cycle is structurally different," or "geopolitical risk is contained," it does so in part because the inputs producing the narrative are no longer independent.

The result is a potential structural blind spot. By construction, anything called a tail risk is something the consensus has decided to look at less directly. The institution that runs a single dashboard with a single methodology is not running a risk system. It is running a confidence machine. And every postmortem cited above, from Wohlstetter forward, identifies a variant of this same structural failure as the proximate cause of the breakdown it is investigating.

The lesson of the literature is therefore not "collect more data" or "hire better analysts." Both of those things have been tried. The lesson is that intelligence, in the operational sense, is the act of noticing when one signal disagrees with another, and refusing to reconcile the disagreement away. That is a method. It is not a person, and it is not a dashboard.

What an organization actually needs, in any high-consequence domain, is an integration layer that is structurally incapable of consensus. Something that surfaces the parts of the picture that disagree with the rest of the picture, and forces the question those gaps imply. Pearl Harbor in 1941, Yom Kippur in 1973, September 2001, Silicon Valley in 2023, and Texas in 2021 are five different events across eight decades, several different domains, and a single forensic profile. The warnings were there. The disagreement was visible. The method for surfacing it was not.

After enough of these, the question stops being whether a method like that is valuable. It evolves into whether an institution is willing to operate without one.

About the Author

John Carmean developed Pretelligence™, a tensor-based multi-domain detection platform with directional classification and probability-weighted response execution. The system is validated across energy, financial, transportation, and consumer domains using U.S. public domain datasets with locked parameters and no domain-specific tuning. Sixteen provisional patents have been filed covering the detection architecture, multi-domain coordination, and response execution systems. Patent applications 63/927,459 through 64/036,916.

Sources

[1] Roberta Wohlstetter, Pearl Harbor: Warning and Decision (Stanford University Press, 1962). Bancroft Prize, 1963.

[2] Joint Congressional Committee on the Investigation of the Pearl Harbor Attack, Hearings, 39 volumes (U.S. Government Printing Office, 1946).

[3] Agranat Commission, Interim Report on the Yom Kippur War (Government of Israel, April 1, 1974).

[4] Uri Bar-Joseph, The Watchman Fell Asleep: The Surprise of Yom Kippur and Its Sources (State University of New York Press, 2005).

[5] National Commission on Terrorist Attacks Upon the United States, The 9/11 Commission Report (U.S. Government Printing Office, July 2004), particularly Chapter 8, "The System Was Blinking Red."

[6] President's Daily Brief, Bin Ladin Determined to Strike in US, August 6, 2001, declassified April 10, 2004.

[7] Federal Bureau of Investigation, Phoenix Field Office (Kenneth Williams), Electronic Communication, July 10, 2001.

[8] Board of Governors of the Federal Reserve System, Review of the Federal Reserve's Supervision and Regulation of Silicon Valley Bank (April 28, 2023), under signature of Vice Chair for Supervision Michael S. Barr.

[9] Federal Energy Regulatory Commission and North American Electric Reliability Corporation, The February 2021 Cold Weather Outages in Texas and the South Central United States (November 16, 2021).

[10] Constance Babington Smith, Evidence in Camera: The Story of Photographic Intelligence in World War II (Chatto and Windus, 1958).

[11] F. H. Hinsley et al., British Intelligence in the Second World War: Its Influence on Strategy and Operations, 5 volumes (Her Majesty's Stationery Office, 1979 to 1990).