DoctrineDesk by Sweent: Our Capability Demonstration at SHINE 26

On 8 October 2026, Julian Tejera of Sweent LLC presented DoctrineDesk by Sweent at SHINE 26, part of the NATO Innovation Continuum 2026, as an invited speaker. Sweent was selected from among hundreds of companies and invited to take part in the NATO Innovation Continuum 2026, and to present at SHINE 26. The recording of that presentation is above, with English captions.

DoctrineDesk is a capability demonstration: a working AI research and planning assistant that answers only from an approved library, cites every sentence it writes, refuses what it cannot support, and keeps a record of everything it does. We built it to show what Sweent's governed-AI engineering can do. This article walks through what we showed on stage and how it works underneath.

DoctrineDesk is an internal codename for a capability demonstration by Sweent LLC. It is not a commercial product and is not offered for sale. It was developed and funded entirely by Sweent LLC and is not sponsored or endorsed by NATO or any NATO body. All scenarios shown are fictitious. Parts of this recording were blurred or removed for confidentiality and privacy.

The problem we set out to solve

Planning staff work from a large body of published doctrine and research. They need to find what it says, compare sources, and turn the result into documents that other people will check.

A general-purpose AI model looks like a natural fit for that work, and it is a poor one. It answers from whatever it absorbed in training. It cannot show where a claim came from. When it does not know, it tends to produce a confident answer anyway. For someone whose work will be reviewed line by line, an answer with no source behind it cannot be used.

We built DoctrineDesk the other way around. The assistant has a fixed library and nothing else to draw on. Before it writes anything, it retrieves the passages that bear on the question, and it may only answer from those passages. Around that sits a set of checks that every request has to pass. Each check fails closed: if it cannot be completed, the request is refused. All of it is written to an audit record.

A library you can see

The demonstration started with the library, because the library is the boundary of what the assistant is allowed to say. At the time of the demonstration it held 466 sources split into 23,423 passages: NATO doctrine, NATO Defense College research papers and a fictitious exercise scenario. All of it was unclassified, publicly released material.

That boundary was visible on screen. A planner could open the list of indexed sources and read the passages a question would be answered from. Nothing the assistant knows is hidden from the person using it.

Asking a question and checking the answer

We asked a question in plain language about the "Zoran Sea" exercise scenario, which is fictitious. The answer was assembled only from passages in the indexed library, and each sentence carried its source and page. A reader who doubted a sentence could go straight to the passage behind it.

The same exchange produced a chart to go with the answer, and the result was saved as a document to come back to later.

What it refused, and why that matters

An assistant like this is judged as much by what it declines as by what it answers. We showed three refusals on stage.

1. A question outside the library. The assistant refused it as out of scope instead of guessing. If the sources do not contain the answer, there is nothing to cite, so there is nothing to say.
2. A prompt-injection attempt. A request written to make the assistant ignore its instructions was stopped before any model was called.
3. A targeting and execution request. This was refused as well. The demonstration is for the planning side only.

Each refusal was recorded with its reason. Someone reviewing the work later can see what was asked, that it was refused, and why.

Research and red team

Some questions need more than one lookup. In research mode the assistant ran its own searches across the literature, showed each search as it made it, and corroborated its answer across several sources before presenting it. The planner saw how the answer was reached as well as the answer.

Red-team mode works from a single instruction: "Tell me why I'm wrong." A planner stated an intent, and the assistant set the literature against each assumption behind it. The result was labelled as an academic perspective. It does not make the decision. It puts the strongest published objections in front of the person who will.

From a question to planning documents

The planning part of the demonstration took a scenario through to drafts a staff could work on: a strategic assessment, candidate response options, and the documents that go with them. These were produced as Word, Excel, PDF and PowerPoint files. Each one stayed a draft until a person released it.

We also showed doctrine as a map. A planning doctrine was drawn as stages and steps, and each step rested on a quote checked against the source document. That part of the recording is blurred.

An audit trail for everything

Every exchange in the demonstration was recorded: the question, the passages retrieved for it, and the verdict of every gate it passed through. Any of them could be replayed afterwards. The overview page drew its totals from that same record, so the numbers on screen and the evidence behind them could not drift apart.

This is the part of the design we consider most important. It is what lets a reviewer trust the assistant's output without having to trust the assistant.

What this shows about how we build

DoctrineDesk is a retrieval system over an approved library, a language model (Anthropic Claude) that may only answer from what was retrieved, a set of fail-closed checks around every request, and an audit record of all of it. Sweent LLC designed, built and paid for all of it.

It is not a product. It is not for sale, and no price, licence, release date or roadmap exists. What it shows is the engineering: retrieval a person can inspect, a model held to its sources, checks that refuse when they cannot verify, and a record that accounts for every answer. That is how we think AI should be built wherever the answer has to be defended.

At SHINE 26

The event exists for this: the working groups presented their solutions to delegates, NATO Allies and affiliates. Besides the stage presentation recorded here, the organisers rotated groups of attendees through the working groups, and Sweent gave short presentations of DoctrineDesk to each group in turn.

We wrote about being selected earlier in the year, in our post on the NATO Innovation Continuum 2026.

About the recording

The video is an edited recording of a live presentation, not raw footage. Some material that was on screen was blurred for confidentiality, and the picture was cropped and cut so that attendees' faces and other participants' identifiable intellectual property do not appear. Where the presenter shows the application, the screen recording is used in place of the camera's view of the stage screen. Waiting time while the assistant worked was shortened; the presenter says so each time, and the walkthrough labels it. Background noise was reduced, and the captions were machine-transcribed and then edited for spelling, names and clarity.

Disclosures

Not a product. DoctrineDesk is a capability demonstration. It is not a commercial product or service, it is not offered for sale or licence, and Sweent LLC makes no commitment that it will be released. Nothing in this video is an offer, a quotation, or a statement of future availability.

The name. "DoctrineDesk" is an internal codename used to refer to this demonstration. Sweent LLC does not claim it as a trademark or service mark.

Independently developed and funded. DoctrineDesk was conceived, designed, developed and funded entirely by Sweent LLC, at its own expense. It was not sponsored, funded or paid for by NATO, by any NATO body or affiliate, by any government, or by any other organisation or event participant. No such party has any ownership interest in, or claim to, the DoctrineDesk software or its design, which remain the sole property of Sweent LLC.

Participation, not endorsement. Sweent LLC was selected from among hundreds of companies and invited to take part in the NATO Innovation Continuum 2026. It was invited to build a solution in one of the event's working groups and to present it at SHINE 26, where the working groups showed their solutions to delegates, NATO Allies and affiliates, and where Julian Tejera spoke as an invited speaker. NATO does not endorse DoctrineDesk, Sweent LLC, or any Sweent product or service, and selection, participation and the use of publicly released NATO publications as source material do not imply endorsement, approval or accreditation. Any NATO name or emblem visible in the recording is part of the event venue as filmed.

Fictitious scenarios. Every scenario, question and plan shown is fictitious or hypothetical and was created only to demonstrate capability. The "Zoran Sea" material is an exercise scenario, not a real place or event. Questions that mention real countries are illustrative research prompts; they are not plans, advice, assessments or positions of Sweent LLC, NATO, or any government.

Not for operational use. The demonstration is a research and drafting aid over unclassified, publicly released material. Its output is machine-generated and must be checked by a qualified person. It handles no classified information.

Source material. The NATO Defense College releases its publications free of charge under its published terms for external use, which rely on fair use. Sweent used them as the demonstration's library under those terms: each use is credited to its source, quoted briefly and linked to the original. Sweent does not sell, redistribute or republish NATO or NATO Defense College publications, and claims no rights in them; they remain the property of their owners. Views expressed in NATO Defense College publications are those of their authors and do not necessarily reflect the opinions of the NATO Defense College, NATO, or any member government.

Edited recording. This video is an edited recording of a live presentation. Some material was blurred for confidentiality. The picture was cropped and cut so that attendees' faces and other participants' identifiable intellectual property do not appear. Waiting time was shortened, the sound was cleaned, and the captions were edited for clarity.

Other marks. The credits show Sweent LLC's own company credentials: GSA contract holder (#47QRAA25D0024), IBM Partner Plus Silver Partner, and "Built for AWS GovCloud (US)". They describe Sweent as a company, not DoctrineDesk. All third-party names and marks belong to their owners.