CIA’s 1987 Data Evaluation and the Statistical Cost of Covert Ops
By Harlan Pryce ·
Kept secret until 2000 and released as a fragmented wreck, the 1987 data evaluation shows the CIA attempting to turn the unpredictability of human assets into a mathematical cluster.
The assets of the late Cold War were not people so much as they were liabilities waiting for a mathematical trigger. In the covert wars of the 1980s, a misplaced trust in a foreign national didn't just end a career; it ended lives in the jungles of Central America and the streets of Kabul.
The record survives here as a series of OCR-extracted fragments, more a digital ruin than a readable report. It is a ghost of a file, an artifact released in August 2000 that has been stripped of its coherence by the very process of its digitization. What remains are the skeletal remains of a project titled "1987 SUMMARY DATA EVALUATION - OPS," along with a scatter of alphanumeric noise: "S ~MM 1 1 /2,17,7 EV,i4 4r/vim," and "Svf1Mf 7 i2,l?-~ EV,1&u4r/&/ l q S~7" (translated from the record). These fragments are the only evidence we have of a moment when the CIA's Directorate of Operations—the clandestine arm responsible for the recruitment of foreign assets and the execution of sabotage—decided that the chaos of espionage could be solved with a formula.
The Logic of the Cluster
To understand what the CIA was doing in 1987, one has to understand the tool they were applying to their operational failures. According to the public record, cluster analysis is a data analysis technique aimed at partitioning a set of objects into groups such that objects within the same group exhibit greater similarity to one another than to those in other groups. In the world of statistical analysis, this is a primary task of exploratory data analysis. In the world of the CIA, it was a way to turn human beings into data points.
If you are running "Operational Projects" across a dozen different time zones, you are dealing with a staggering amount of noise. You have assets who betray you, assets who are captured, and assets who deliver gold but lie about the source. The goal of the 1987 evaluation was to find the "cluster" of traits that predicted a successful operation versus the cluster that predicted a catastrophe. The record shows an obsession with "SUMMARY DATA," suggesting that the Agency was no longer looking at the individual stories of their spies, but at the aggregate patterns of their failures.
1987 and the Operational Ghost
The timing of this evaluation is the key. In 1987, the CIA was reeling from the public exposure of the Iran-Contra affair, a scandal that revealed the agency's illegal efforts to fund the Contras—the right-wing rebel groups fighting the Sandinista government in Nicaragua, the Central American nation where the CIA sought to destabilize the socialist regime. The Iran-Contra disaster was an operational failure of the highest order, characterized by a lack of oversight and a total breakdown in the chain of command. It proved that the "gentleman spy" era of intuition and personal relationships was producing results that were both illegal and unsustainable.
This is where the connection becomes clear. The "1987 SUMMARY DATA EVALUATION" was not a routine audit; it was a post-mortem of a broken system. The Agency was attempting to move away from the idiosyncratic judgments of case officers and toward a scientific model of risk. By applying cluster analysis to their "Ops," they were searching for the common denominator in their disasters. They wanted to know if the failures in Nicaragua shared a statistical cluster with the failures in Afghanistan or the botched infiltrations in Eastern Europe.
When the record mentions "SUMMARY - 1987 OVERALL DATA EVALUATION - OPERATIONAL PROJECTS" (translated from the record), it is describing a desperate attempt to quantify the unquantifiable. The CIA was trying to build a map of betrayal. If they could identify the cluster of behaviors, backgrounds, or environments that led to an asset turning double, they could stop the bleeding before the next scandal hit the front page of the New York Times.
The Noise as a Shield
There is a profound irony in the state of this document. The record was released in 2000, but it is virtually unreadable, a jumble of "CBD c7-cu'! S~"VFS.: VU" and "S-CrrvN _-- $C,~Ct'S" (translated from the record). While the Agency often uses black ink to hide names, here the redaction is structural. The data has been rendered into noise. This is not merely a failure of OCR technology; it is the shape of a gap. The Agency released the fact that they performed a data evaluation, but they withheld the results of that evaluation.
If the results had been innocuous, they would be legible. The fact that the specific clusters of operational failure remain hidden suggests that the 1987 evaluation found something the Agency still considers dangerous. Perhaps they discovered that their successes were as random as their failures. Perhaps they found that the most "reliable" cluster of assets was actually the most compromised. The gaps in the text are the places where the math failed to save them.
The Algorithmic Asset
The desk's reading is that this file represents the exact moment the CIA transitioned from human intelligence to predictive intelligence. The 1987 evaluation was the prototype for the modern surveillance state's approach to human behavior. By attempting to "cluster" operational projects, the Agency was admitting that the human element—the intuition of the case officer, the loyalty of the asset—was a variable that needed to be managed by a computer.
This was the birth of the algorithmic asset. The Agency stopped asking why a person betrayed them and started asking which group that person belonged to. If you fit the cluster of a "high-risk asset," you were treated as a liability regardless of your actual performance. The human cost of this transition was a shift in the nature of betrayal; the spy was no longer a partner in a secret war, but a data point in a statistical model.
What a full release of these files would show is the specific criteria the CIA used to group their agents. It would reveal the biases—racial, national, and psychological—that the Agency baked into its mathematical models of loyalty. The still-withheld pages are protecting the logic of the machine, the specific weights and measures used to decide who was a "cluster" for success and who was a target for disposal.
The pattern suggests that the CIA never stopped this project; they simply perfected it. The fragmented noise of the 1987 record is the ancestor of the modern predictive algorithms used to flag "suspicious" behavior in global communications. The Agency learned in 1987 that you don't need to know the man if you know the cluster. And in the end, the people who paid for this shift were the assets themselves, who were reduced to a set of coordinates in a summary data evaluation, only to be discarded when the math decided they no longer fit the pattern.