One Active Investigation

Seven years before a U.S. missile hit the Shajareh Tayyebeh Elementary School in Minab, an intelligence analyst recorded that part of the compound was no longer a naval facility but an elementary school. The finding survived in a government system disconnected from the database used to build the target package.

When officers built the target package for Minab, the analyst's finding was absent from the authoritative targeting database, according to what investigators have since told Congress.

On the first morning of Operation Epic Fury, a Tomahawk missile hit the Shajareh Tayyebeh Elementary School in Minab. Human Rights Watch, the New York Times and a letter from 25 senators put the dead at 175 or more. Amnesty International's own investigation counted 168, then revised the figure down to 156, including 120 children, to match the numbers Iranian authorities were releasing. CNN, citing Iranian state media, reported at least 168 children and 14 teachers. The counts disagree, but every account describes a building full of children and teachers.

The analyst had been right for seven years, in writing, in a government system, and none of it reached the officers who built the target package. They were working from imagery that had not been updated in years, on a compound where a fence had gone up in 2016 to separate the school from the base, and where satellite images from December 2025 showed dozens of people apparently playing in a courtyard.

The analyst had identified the school seven years before the strike, but the finding remained trapped in a system the targeting officers did not use.

What a warning is worth if you can waive it

The targeting system carried a second warning.

Messages sat in the targeting system flagging that intelligence on many Iranian sites was years old and needed re-vetting. The flags were not decorative. Clearing one required a senior officer to sign off before a site could be added to the strike list. Three sources told CNN that senior commanders bypassed those warnings for what two of them called "expediency," in a rush to provide targets at the start of the war.

Intelligence on many of the sites added to the list was over ten years old. Analysts had triaged, refreshing records for "upper-tier" targets first, the ones most likely to be hit early. They did not get to all of them before the shooting started.

"The Pentagon is pressing everyone to move faster," one source told CNN. CENTCOM leadership, the source added, "did not push back either."

The override did not bypass the process. It used an exception the process allowed: stale intelligence could move onto the strike list if a senior officer signed. Under pressure to produce targets quickly, the safeguard depended on the same leadership setting the pace.

The office that would have argued

There was, on paper, a third safeguard: people whose entire job was to sit inside target development and raise a hand about civilians.

The Civilian Harm Mitigation and Response program grew out of two decades of American air wars in which Airwars estimated more than 20,000 civilians were killed across Iraq, Syria, Afghanistan, Libya, Pakistan, Somalia and Yemen. In 2022 the Pentagon published an action plan, and around 200 personnel were eventually assigned to the work, embedded in planning processes and the targeting cycle.

Then it was taken apart. Staffing fell to roughly a tenth of its original size. Several combatant commands lost their civilian harm teams entirely. CENTCOM's went from about ten people to one, which is where it stood during a campaign that expended more than thirteen thousand munitions. The Civilian Protection Center of Excellence dropped from about 40 staff to nine. Congress had appropriated $11.6 million for the work against a departmental estimate that it needed $72.8 million, and the following year's budget request reportedly zeroed the line out entirely.

The Under Secretary of Defense for Policy and the Secretary of the Army each asked Secretary Pete Hegseth to eliminate the Center of Excellence outright and rescind the civilian harm policy. Congress had written both into statute, leaving them harder to eliminate than a budget line.

The Defense Department's own Inspector General reported in May that the department "did not fully implement any of the CHMR-AP objectives" and now lacked "many of the personnel and tools designed to execute" its own civilian harm policy. On the day that report landed, Representative Adam Smith told the Secretary of the Army: "You are in violation of the law right now on civilian harm."

Among the casualties, according to the Center for Civilians in Conflict, was the red team, the element whose assigned function was to challenge assumptions and flag civilian risk before a strike. Losing it, the organization noted, "has weakened a critical internal check for preventing confirmation bias."

The people whose job was to disagree had been removed from the room. The investigation into Minab then had to explain what happened without them.

Thirteen thousand six hundred twenty-nine

On May 14, Senator Kirsten Gillibrand asked the commander of Central Command about reporting that strikes had hit 22 schools and multiple hospitals in Iran.

"There is no indication that we have, that has been corroborated. Zero," Admiral Brad Cooper said.

"Then how many schools did we bomb?"

"There is one active civilian casualty investigation from the 13,629 munitions."

"So, how do you explain the publicly available information that 22 schools have been hit and multiple hospitals?"

Cooper said the open-source numbers he had seen were not accurate. Pressed further, he acknowledged that the military had not investigated the incidents in question. He said he would work on a report about whether hospitals or schools had been destroyed in U.S. strikes.

Five days later he told the House Armed Services Committee that he had spent the weekend with his team conducting exactly those assessments, and had determined that only the Minab school correlated with an American strike and American munitions.

That finding may well be correct. But the organization that fired the munitions spent a weekend reviewing its own munitions and reported back that it had one active civilian-harm investigation. The process offered no outside check on the command's conclusion.

"Zero corroborated" meant only that the command's own process had not confirmed harm, after Cooper acknowledged that it had not investigated those incidents. The Pentagon treated that missing evidence as proof that nothing had happened.

Airwars, meanwhile, has documented hundreds of civilian casualty incidents in Iran. The World Health Organization put the death toll at 3,375 before the ceasefire collapsed in July. Asked for an update that same month, CENTCOM told The Intercept: "We have no new updates at this time pertaining to current investigations or munitions that were expended during Operation Epic Fury."

How other fields separate the checker

When an airliner goes down, Boeing does not write the accident report. The National Transportation Safety Board does, and Congress built it as a separate agency with its own appropriation and no regulatory authority at all, precisely so it would have nothing to protect by shading a finding. When a drug trial starts hurting people, the company that makes the drug does not decide whether to stop it; a Data Safety Monitoring Board sees the unblinded data while the sponsor stays blind to it. The counter-case is documented just as carefully. Japan's nuclear regulator sat inside the ministry promoting nuclear power, and after Fukushima the Diet's investigation commission called the arrangement regulatory capture.

NASA's Independent Verification and Validation program exists because the contractor writing Shuttle software was also the party verifying it. The fix was independence along three axes at once, technical, managerial, and financial. The IV&V team draws its budget from the Office of the Chief Engineer rather than from the project it examines, picks which components to scrutinize, and sets its own schedule. The United States government concluded, in writing, that self-validation of safety-critical software was unacceptable. It reached that conclusion about code that carries astronauts, and has not reached it about software that nominates targets.

AI systems have reached the same conclusion from the technical side. DeepMind's Aletheia math agent used a separate verification pass to catch high-confidence errors that the generation process could not reliably identify in its own output. The generator and verifier had to be separated because asking the same process to check itself left convincing mistakes intact.

Against that standard, the American kill chain offers limited independence. Target validation boards sit under the commander whose targets they validate, and civilian harm assessors report to the commander whose strike they assess. Open doctrine does require that anyone assessing civilian harm must not have been directly involved in the event, but the review remains post-strike, inside the same command, and is currently administered by an office cut to a tenth of its size.

The part that gets automated next

In April, the Civilian Protection Center of Excellence completed its annual report. The Pentagon said it would be posted publicly. It was not. Nick Turse obtained a copy.

It describes an office reduced to nine people. Its civilian harm records database had been cancelled mid-development. Current and former officials told The Intercept that the remaining staff were prevented from investigating major civilian-casualty incidents "like" the Minab school strike. The experts on civilian harm had been pushed away from the cases that most required their scrutiny.

The report describes two distinct AI efforts. First, the Center "leveraged the rapid adoption of the AI-enabled Maven Smart System" to develop a tool giving field commands a fuller picture of the civilian environment. Maven is Palantir's operational platform, used across targeting and command workflows.

Separately, the Center equipped AskSage, which the report calls the Army's enterprise AI platform, with civilian-harm mitigation and response data. The resulting AI assistant prototype was designed to help staff analyze the civilian environment during operations. In an Army training exercise, that assistant cut the time required to produce civilian assessments by 97 percent.

In May, the Pentagon's Chief Digital and AI Officer credited Maven together with the War Data Platform with enabling U.S. forces to strike upwards of 13,000 targets and compress targeting timelines from days into seconds. Palantir's documentation describes its targeting workflow as carrying targets from identification through execution and evaluation, with battle-damage assessment included as one stage.

Joint targeting doctrine divides the work into six phases, from the commander's initial guidance through target development, validation, execution and combat assessment. Those divisions are supposed to preserve different kinds of judgment at different moments. In Project Maven, Katrina Manson describes a redesign at the 18th Airborne Corps in which a human remained "in the loop" at two points. Writing about that redesign alongside the doctrine, she says Maven was "plucking humans out from their traditional role in the cycle in four of six places."

Manson's account relies heavily on interviews with participants, while many operational details remain classified. The Pentagon's public description points in the same direction — compressing a process from days into seconds necessarily changes how much time each step can contain.

Vice Adm. Frank Whitworth, then director of the National Geospatial-Intelligence Agency, focused on what that compression could erase. Told that an operator might click a dot and send a target to a weapons system, he asked, according to Manson: "Tell me about the recordkeeping?" A human decision can be reviewed only if the evidence, changes and objections that produced it survive the decision.

C. Anthony Pfaff, a retired Army colonel who studies military ethics at the Army War College, identified a second problem. A clean digital recommendation, produced by a system the reviewer cannot fully inspect, can make disagreement harder rather than easier. "That's how you get automation bias," Manson quotes him. "It's not hard to get humans to trust machines."

The Pentagon says Maven does not displace human authority. Whitworth has described it as "a decision-making aid" and "not a decision-maker". Accountability, he said, "is still with a commander." Whether that authority remains meaningful depends on operational details the Pentagon has not systematically disclosed: how many targets each reviewer handles, how much time they receive, what evidence they can inspect, and what happens when they slow the process down.

Whitworth's recordkeeping question has a partial public answer. In 2024 NGA launched an accreditation pilot for GEOINT AI models, evaluating the methodology and robustness of model development and testing. The pilot is a real attempt to standardize scrutiny, and it remains inside the agency producing the models. Public materials do not describe an independent body able to reconstruct how a target moved through the system, test the evidence behind an operator's confirmation, and examine what happened after a strike. Classified procedures may exist. No public account establishes that separation.

The public record does not show Maven adjudicating harm caused by Maven-enabled operations. It does show AI-enabled tools being applied to targeting and civilian-harm assessment while the Civilian Protection Center of Excellence was cut by roughly 78 percent and the database where its findings would have accumulated was cancelled.

Wes Bryant used to be the Center's chief of civilian harm assessments. He was pushed out last year after speaking publicly about the dismantling.

"The Hegseth regime held on to a thread of the program mostly in effort to be able to say they are following the law," he said, "but the reality is the Center and its efforts have had little to no influence on or oversight or visibility on operations." The staff who remain, he said, are "supposed to shut up, keep their heads down, and pretend that the civilian harm mitigation and response mission isn't simply a shell for legal CYA."

A currently serving defense official was asked to assess that characterization. The reply, in full: "Yup. No notes."

The rate at which nobody knows

No public evidence shows that an algorithm chose the school in Minab. The investigation that might settle what happened was submitted to senior officials in April and had not been released as of July, when 25 senators wrote to demand it. They asked for an unclassified version suitable for release and a plan to prevent recurrence. Their letter focused on "the integrity of U.S. target development, the adequacy of target validation and vetting procedures, the interoperability of intelligence and targeting databases."

The analyst's 2019 note was exactly the kind of record that poor interoperability kept from the targeting system.

The Pentagon's answer to concerns about military AI is that a human commander remains responsible. But "human in the loop" says nothing about the conditions under which that person decides. A reviewer handling a few targets, able to inspect the underlying imagery and return one for more work, is exercising judgment. A reviewer handling hundreds, unable to inspect the system's evidence and pressed to keep pace, may be performing a step. Both arrangements leave a human in the loop. Only one reliably preserves a check.

The Pentagon has not systematically disclosed how many targets each reviewer handles, how much time they receive, or what they are allowed to see and slow down. It does say its systems can compress targeting from days into seconds. The old process contained hours, handoffs and manual work during which a stale record might still be caught. That friction was never a designed safeguard. It was time in which a safeguard could operate.

Cooper's testimony left the committee with the institutional record: 13,629 munitions expended, an initial review completed over a weekend by the command that carried out the campaign, and one active civilian-harm investigation.