Pentagon review ties Minab school strike to Maven reliance and stale data
Bloomberg reports that an unreleased internal Pentagon review found Central Command staff expected Palantir's Maven Smart System to catch outdated intelligence, while the team meant to check targets for civilian harm had shrunk from about 10 people to one.
The school was still filed as a military site in old records. No civilian harm reviewer checked it before February 28.
Bloomberg reported on September 20, 2026 that an unreleased internal Pentagon review links the February 28 strike on Shajarah Tayyebeh Elementary School in Minab, southern Iran, to three things: outdated target records, heavy reliance on Palantir's Maven Smart System, and deep cuts to the staff who check targets for civilian harm. The findings come from officials familiar with the review, which has not been published.
Two Tomahawk missiles hit the school on the opening day of the US war with Iran. More than 150 people were killed, at least 123 of them children, by Bloomberg's count, Gizmodo reported. The UN has put the number of children killed at 120, IBTimes UK reported.
A record that stopped matching the ground
According to Bloomberg, the site was catalogued as an Islamic Revolutionary Guard Corps facility in outdated data. When it went into Maven with other candidate targets, it came out as a recommended day-one target. The school sat fewer than 100 yards from an IRGC naval installation, and Amnesty International found it had been inside the compound perimeter until a wall went up between 2013 and 2016, IBTimes UK reported.
Evidence of the change existed. Gizmodo, summarizing Bloomberg, says a 2018 satellite image shows a soccer pitch, assembly rows, and playground markings. One analyst noticed changes as early as 2019 and logged them in a system that was not connected to the main intelligence database used for targeting.
Expecting Maven to flag stale records
Officials told Bloomberg that some Central Command staff expected Maven to flag stale records or contradictions in the intelligence assembled for targets. Gizmodo notes it is not clear why they thought the system would do that. Maven fuses satellite imagery, drone feeds, and signals intelligence into strike recommendations under a Pentagon contract worth up to $1.3 billion, IBTimes UK reported, and Yahoo News says it cut target-list work from hours to minutes.
Palantir told Bloomberg it "is not responsible for the underlying data nor identifying intelligence deficiencies" and that there is no evidence its software was at fault.
One reviewer left for all of Central Command
Staffing on Pentagon civilian harm mitigation teams fell by roughly 90% in the years before the war, to fewer than 20 people across the department, Bloomberg reported. Central Command's team went from about 10 people to one. No member of a civilian harm mitigation team reviewed the Minab site before the strike.
On September 22, Bloomberg reported the changes that followed: a revised workflow for refreshing and vetting planned targets, new open-source feeds to track civilian movement, and dozens of Maven upgrades. One of them lets Maven re-review underlying intelligence and flag inconsistencies that human review may have missed.
What the reporting does not say
The review is not public, and every finding above comes from unnamed officials speaking to Bloomberg. A Pentagon spokesperson said the incident remains under investigation and declined further comment, according to Gizmodo. The reporting describes several failures acting together. It does not assign the outcome to any one of them.
Why a build studio cares
We build AI workflows and agents that rank, filter, and recommend, and this report reads like a list of the failure modes we design against. First, the age of a record is data. Every input to a ranking should carry when it was last confirmed and where it came from, and an old input should show up in the output instead of disappearing into a confident score. Second, write down what the tool checks and what it does not, and put that list where operators see it. By Bloomberg's account, users trusted Maven to catch a class of error it was not set up to catch, and the check they assumed is now being added as an upgrade. Third, a review step staffed by one person for a whole command exists mostly on paper. If a design depends on a human check, measure whether the check runs, and make the system report when it did not.
Next step: read Bloomberg's investigation, the primary source for the review's findings.