Against Retrospective Certainty: Temporal Honesty and the Governance of Institutional Memory

Every organization, from a government intelligence agency to a technology company, makes decisions based on what it knows at the moment. These decisions are often made in a fog of uncertainty, where the outcome is unknown and the evidence is incomplete. Years later, when the results are clear, people look back at those decisions and often see them as having been more confident and certain than they actually were. This is a common human tendency known as hindsight bias, where the brain rewrites the past to make the future seem inevitable. However, in large institutions, this rewriting is not just a trick of the mind; it is also a problem of paperwork. When the original documents, notes, and doubts are lost, overwritten, or separated from the final conclusion, the history of how a decision was made becomes impossible to check. Without the original record, an institution can accidentally, or even intentionally, present a past judgment as a solid fact when it was actually a guess, a worry, or a minority opinion.

A new paper by researcher Ilias Konstantinidis tackles this problem by proposing a standard called "temporal honesty." This concept suggests that institutions have a duty to keep the messy, uncertain, and evolving nature of their past thinking intact. It is not enough to save the final report; the organization must also preserve the doubts, the disagreements, the assumptions, and the specific evidence that existed at the time the decision was made. The goal is to ensure that if someone looks back years later, they can reconstruct exactly what the institution knew and felt at that specific moment, rather than having to guess based on a polished, simplified version of history. The paper argues that without this kind of honest record-keeping, it becomes impossible to hold institutions truly accountable for their past actions, because the evidence needed to test their claims has been erased by time.

To understand why this matters, consider how institutions usually work. They gather information, debate it, and eventually issue a judgment. Over time, as new information arrives or leadership changes, that judgment is updated. In a system without temporal honesty, the old version of the judgment might simply be deleted or replaced by the new one. The original reasoning, the specific data points that were used, and the voices that argued against the final decision disappear. This creates a gap between what actually happened and what the surviving documents say happened. The paper calls this gap "retrospective certainty," a state where the past looks cleaner and more unified than it really was. The author suggests that to fix this, institutions need to treat their past states of knowledge as permanent records that cannot be silently overwritten. Instead of deleting an old assessment, they should link it to the new one, showing clearly what changed, what was added, and what was removed.

The paper tests this idea by looking at two very different real-world examples where the ability to reconstruct the past was crucial. The first case involves the investigations into weapons of mass destruction in Iraq. Several official inquiries were launched to understand how intelligence agencies reached the conclusion that Iraq possessed such weapons. The researchers found that these inquiries struggled because the original records were incomplete. In one instance, a review committee thought it had seen all the relevant intelligence papers, only to discover later that eight important documents had been left out. In another case, the committee could not find the original drafts of a key report, only copies that had been altered. Most importantly, the inquiries found that dissenting opinions—where experts had warned that the evidence was weak—were not properly recorded or were hidden from the people who needed to see them. Because these records were missing or obscured, the investigators had to spend years trying to piece together what the original situation looked like, and even then, some details remained unclear. The paper uses this to show that when the "lineage" of a decision is broken, it becomes very hard to know if the institution made a mistake or if the record itself was flawed.

The second example looks at how artificial intelligence companies document their work. As these companies release new versions of their AI models, they publish reports explaining how the models perform. The researchers examined how these reports changed over time. They noticed that while companies kept talking about the same types of risks, like the models making up facts, the specific tests they used to measure those risks kept changing. A test used in one year might be replaced by a different test the next year, with no clear note explaining how the two related. For an outside observer trying to track whether the AI was getting better or worse, this lack of a clear connection made it difficult to compare the past with the present. The paper argues that even in the public eye, without a clear chain showing how one evaluation led to the next, it is impossible to know if a risk has been truly addressed or just ignored because the measurement tool changed.

The core finding of the paper is that preserving the "epistemic state"—a fancy way of saying the total picture of what was known, doubted, and decided at a specific time—is a distinct and necessary part of good governance. It is separate from whether the decision itself was right or wrong. An institution can make a terrible decision based on bad data, but if the record of that bad data and the bad decision is preserved honestly, people can learn from it. If that record is lost or smoothed over, the lesson is lost too. The author proposes a set of practical rules to achieve this. These rules include keeping a version history so that changes are visible, making sure that uncertainty and disagreement are attached to the final conclusion rather than stripped away, and keeping a record of who saw what information and when. The paper acknowledges that some information must remain secret for safety or privacy reasons, but it argues that even in those cases, the structure of the record—the fact that a decision was made, who made it, and that it was based on certain types of evidence—should remain visible.

This approach does not solve every problem of organizational failure. It cannot force leaders to listen to warnings or stop them from making bad choices. It cannot fix a culture where people are afraid to speak up. However, it does ensure that when things go wrong, the truth is not buried under layers of rewritten history. By demanding that institutions keep their past selves visible, complete with all their doubts and errors, the paper offers a way to make the future more honest. It suggests that the best way to learn from the past is not to remember it perfectly, but to keep the original, messy evidence available so that it can be examined exactly as it was, without the distortion of knowing how the story ended.