When Search Becomes Inquiry

What the Palme investigation archive reveals about the future of archival access

On February 28, 1986, Swedish Prime Minister Olof Palme was shot on a Stockholm street on his way home from a cinema. No one has ever been convicted. The investigation that followed became one of the largest criminal inquiries in Swedish history, and by the time prosecutors formally closed it in June 2020, it had produced an archive so large that Swedish police have compared its scale to Lockerbie and the Kennedy assassination: something on the order of 250 meters of shelving, accumulated over three and a half decades.

Forty years after the murder, that archive has a new kind of researcher. A small team of journalists and engineers has built an AI system that does not simply search the Palme files. It reasons across them, generating its own follow-up questions, chaining fragments of testimony and forensic reporting into structured lines of inquiry, and returning answers rather than documents. The tool is called PalmeNet-Chat, and it is worth examining closely, not for what it says about who killed Olof Palme, but for what it says about where archival access is heading once an archive gets too large for any one person to read.

This is not primarily a story about AI solving a decades-old murder. It is a story about what happens when an archive grows past the point where any single reader could get through it, and a system is built to work across it on its own initiative rather than simply retrieving what it is asked for. That shift, from availability, to retrieval, to something closer to reasoning, is the subject of this dispatch, and the Palme case is simply where it happens to be visible first.

The access problem before the AI

Long before any AI entered the picture, the Palme archive already posed an ordinary archival problem at an extraordinary scale. A 1999 Swedish state commission estimated that reading the archive as it then existed would take one person six to seven years of continuous work; by the time the investigation closed in June 2020, police put the figure at closer to nine years for the archive’s full, grown size. Police digitized their internal holdings around the time of that closure. Digitization, though, is not the same as public access. Material can only be released to researchers after a separate, slower confidentiality review, and by most accounts only a small fraction of the archive, on the order of a tenth, has cleared that review and become available at all. What has been released runs to thousands of individual documents, each tens to hundreds of pages, spanning interview transcripts, forensic reports, maps, sketches, and internal police correspondence, most of it in Swedish and none of it originally structured for computational access.

This is a familiar shape of problem for anyone who works with large institutional archives, and its familiarity is the point. The Palme archive did not become hard to access because of its subject matter. It became hard to access for the same structural reasons any sufficiently large, sufficiently old, sufficiently heterogeneous collection becomes hard to access: material accreted over decades under different recordkeeping practices, released to researchers slowly and partially even after digitization, and never arranged with a future computational user in mind. What makes the case unusual is not the underlying access problem. It is who decided to solve it, and how.

What was built, in two stages

The first version of the tool grew directly out of that access problem. Investigative journalists Martin Johnson and Anton Berg, working on the podcast Spår, partnered with the Swedish engineering firm Softwerk to build an on-premise environment that let them query the digitized case files directly rather than searching them document by document. The emphasis fell on retrieval and confidentiality: the material stayed off any external cloud service, open-source models ran on Softwerk’s own servers rather than a commercial provider’s, and the system’s job was to make an otherwise unreadable volume of scanned, largely Swedish-language material searchable at all. This is the kind of infrastructure the field would recognize immediately. It solves an availability problem. It does not, on its own, reason about the material; it finds documents and hands them to a human reader.

The second version, built in a four-day engagement in late 2025 between Softwerk and the AI consultancy ML6, changes the shape of the tool rather than simply its scale. The system moved to cloud infrastructure, adopted Google’s Gemini models in place of the original open-source setup, and gained a hybrid retrieval layer combining keyword and semantic search, along with three selectable modes of operation: an Agentic RAG mode that returns retrieved documents directly, a Deep Research mode that dispatches separate sub-agents to pursue different angles of a question in parallel, and an Agentic Newsroom mode built to mirror the journalists’ own investigative workflow. The distinction between the two versions is not merely that the newer one is faster or covers more documents. It is that the newer one is built to generate its own next question. In one example from the system’s own developers, the system appears to have located a reference to a 1986 interview during an initial search, then issued a follow-up query built around that year on its own, pursuing a lead it had not been explicitly asked to pursue. Even the engineers describing the example are inferring the system’s reasoning after the fact rather than asserting it with certainty, which is itself a small illustration of the transparency problem this dispatch returns to later.

That behavior, a system extending a bounded line of inquiry on its own initiative rather than waiting for the next prompt, marks the point where this stops being a retrieval tool in the traditional sense and starts being something closer to a research assistant working from its own provisional hypotheses.

The methodological shift

Set beside conventional digital forensics, the change in method is easier to name than the change in outcome. Traditional forensic work on a document archive is fundamentally an availability problem: recover the material, index it, make it searchable, and let a human investigator decide what it means. The endpoint is retrieval. PalmeNet-Chat’s second version was built to move past that endpoint. Its documented behavior, issuing an unprompted follow-up query once an initial search surfaces a promising lead, is not faster retrieval. It is the system pursuing a bounded line of inquiry of its own, on the basis of what it has already found.

The endpoint is no longer retrieval. It is evaluating what a system does with material once it has already been found.

That shift shows up again in how the tool is structured to be used. Rather than offering one search mode, it offers three: a mode that returns documents directly, a mode that dispatches parallel sub-agents against different angles of a single question, and a mode built to mirror the way the journalists themselves work through a lead. Conventional forensic tooling generally hands an investigator one lens at a time, full-text search, then metadata filtering, then manual cross-referencing. This tool is built to hold several investigative postures at once and let the user choose which one to apply to a given question.

The clearest statement of what the shift is for, though, comes from Mattias Davidsson, a senior lecturer at Linnaeus University who has spent recent years working with the archive independently of this specific tool. Asked what interests him most about applying AI to the material, he has said it is less the retrieval itself than the possibility of turning raw evidence into structured argument, and then testing how well that argument holds up against competing explanations. That is a different function than forensic search has traditionally served. It is not “here is the document.” It is closer to “here is a claim, and here is how well the archive supports or undermines it.”

None of this changes what the underlying documents say. It changes what a person, or a system, does with them once they can be searched at all, and it suggests that the harder work ahead in this kind of archival access is no longer confined to finding material. Increasingly, it lies in evaluating what a system does with material once it has already been found. Every argument the system constructs still depends, in the end, on the documentary provenance of the fragments it strings together; the reasoning may be new, but its credibility rests entirely on evidential relationships the archive itself established long before any system read it.

What reasoning obscures

A retrieval system exposes documents. A reasoning system exposes conclusions, and that difference carries a risk the first two versions of this tool did not. A search result can be wrong, but its wrongness is usually visible: the document either supports the query or it does not. A synthesized answer can be wrong in a quieter way, plausible in tone, coherent in structure, and confident in delivery, while omitting a contradiction buried in a document it did not retrieve or weighting one account over another without disclosing that it did so. The further a tool moves from returning documents toward returning explanations, the more it matters that the evidential chain behind each claim stays visible to the person reading it, rather than folded into an answer that reads as settled. On this point PalmeNet-Chat’s own design offers a partial safeguard rather than a solution: its retrieval layer returns document fragments rather than whole files, and a separate tool can fetch the full source document by identifier, which at least makes it possible to trace a claim back to a specific passage.

Whether a user actually does that tracing, rather than accepting the synthesized answer at face value, is a separate question the interface itself cannot answer.

What the shift does not solve

A reasoning layer, however capable, is still bounded by what has been made available to it. By most accounts, only a small fraction of the Palme archive, on the order of a tenth, has cleared confidentiality review and been released for research use at all, regardless of how much has been digitized internally by police. Whatever a system can chain together in argument, it can only chain together from that released fraction, and the shape of what remains withheld is itself unknown. A tool that reasons fluently across ten percent of a corpus can create an impression of comprehensiveness that the underlying coverage does not support, and that gap does not narrow just because the reasoning improves.

The move from an on-premise environment to cloud infrastructure raises a second, more familiar problem. Traditional archival practice has generally treated custody, authenticity, provenance, and access as inseparable: the institution that holds the material and controls the environment it is processed in is also the institution accountable for its integrity. This tool’s own history complicates that assumption. Its first version kept confidential investigative material inside a closed, on-premise environment precisely to preserve that kind of accountable custody. Its second version moved the same material, and the analytical capability built on top of it, onto external cloud infrastructure to gain the reasoning power described above. That move does not necessarily compromise the material. It does mean the analytical capability and the custodial environment no longer sit in the same place, which is a real change from the arrangement digital forensics has generally assumed, not merely a routine infrastructure upgrade.

Neither of these limits is a criticism of the tool as built. They are a reminder that reasoning over an archive and establishing the completeness and custody of that archive are separate problems, and that solving the first does nothing, on its own, to resolve the second.

The Palme investigation is an unusually visible example, but the underlying condition is not unusual at all. Government investigations, regulatory and compliance repositories, environmental monitoring records, and ordinary institutional email accumulation all reach, sooner or later, a scale that exceeds what any working professional can read in full. Records managers and digital preservation specialists have generally faced that condition as a retrieval problem, one that better search and better description could eventually solve. This case suggests the more durable version of the problem sits one step further along: not simply whether material can be found, but what kind of reasoning over that material an institution is prepared to stand behind once it has been found.

Closing

The tool built around the Palme archive will not settle who killed Olof Palme, and whether it produces a genuinely new lead is a separate question from the one this dispatch has tried to answer. What it demonstrates, regardless of where the investigation eventually lands, is a shift that extends well past this one case. Archives of comparable scale and fragmentation exist inside government agencies, courts, and institutions well outside the world of criminal investigation, and most of them will eventually face some version of the same choice this project has already made: whether the tool built to make them accessible should stop at finding documents, or should be trusted to reason about what they mean.