The Diplomatic Simulator

The Diplomatic Simulator replays multi-party diplomatic negotiations in which every national delegation is played by an artificial intelligence agent, that is, a language program instructed to speak and reason in the voice of one government. It is a teaching and exploration tool. It is not a forecasting system, and its authors state plainly that it must not be used to inform real policy, negotiation, or intelligence judgements.

Four scenarios have been simulated to date: the Arctic, with seven delegations; Central Asia and the Fergana Valley, with seven; Cyprus reunification, with six; and the South China Sea, with nine. An earlier two-party session on the Strait of Hormuz, involving Iran and the United States, is preserved as the original demonstration. Across the four multi-party scenarios, twenty-nine delegation agents produced ninety-four statements of record.

The organising principle of the design is information isolation. When a delegation agent writes its statement for a round, it receives only its own confidential brief, a neutral public brief that everyone shares, and the public transcript of what has already been said aloud. It never sees another government's private instructions, and it never sees another delegation's private reasoning. This mirrors the informational structure of real diplomacy, where the public record is common property and the mandate is not.

The scenario material is not invented. The confidential briefs are adapted from crisis-negotiation exercise packs associated with the United States Army War College International Strategic Crisis Negotiation Exercise programme, which are taken to universities. The file names preserved in the repository record where each pack was used, including New York University, Penn, Syracuse, and the University of Minnesota.

After the talks conclude, a separate analyst agent reads the whole transcript and produces a scoreboard, a debrief for each delegation, and a neutral summary written in the voice of a convening envoy. A further layer re-runs each scenario eight times under randomly varied external conditions, in order to show which outcomes are stable and which depend on luck.

The repository is candid that every number on the site is one artificial intelligence system's subjective judgement of another artificial intelligence system's writing. There is no ground truth anywhere in the tool. The most serious structural limitation is that a single underlying model plays every side of every table, so the adversaries are not genuinely independent minds.

The problem. Crisis-negotiation exercises are among the most effective ways to teach diplomacy, because they force a participant to argue a position they may not hold, under time pressure, against people who have been told to resist them. But an exercise consumes an enormous amount of scarce human attention. A seven-party Arctic exercise needs seven teams, seven mentors, and two days. It can be run perhaps once a year at any given institution.

The second problem is that exercises leave almost no durable record. What a delegation said in round two, why it said it, and how that changed the shape of the room afterwards is rarely written down in a form anyone can study later. The learning is real but it is locked inside the experience of having been there.

The gap. Existing computational work on negotiation has largely concentrated on games with fixed rules. The best known example is CICERO, the system developed by Meta's Fundamental AI Research team and reported in Science in 2022, which reached human-level play in the board game Diplomacy by combining a language model with strategic planning. That is a considerable achievement, but a board game has a scoring function and a win condition. Real multi-party talks have neither. A separate strand of work has examined what happens when language models are placed in simulated wargames: Rivera and colleagues, publishing at the ACM Conference on Fairness, Accountability, and Transparency in 2024, found that all five models they tested displayed escalation patterns that were difficult to predict. That finding is a caution rather than an endorsement, and this prototype should be read against it.

The response. The Diplomatic Simulator occupies the space between the two. It takes the scenario packs and confidential briefs that human exercises already use, replaces the human delegations with agents, and keeps the entire record. What the tool produces is not a prediction of what governments would do. It is a legible artefact: a transcript, a set of tagged tactics, and a scoreboard, all of which can be read, disputed, and compared against what a human exercise produced from the same brief.

The design commitment that makes this worth doing is transparency about provenance. Every intermediate product of every run is committed to the repository in a readable form: the extracted party profiles, the public brief, the transcript, the analysis. A reader who doubts a score can open the transcript and check what was actually said.

The prototype is designed to reproduce the structure of a multi-party crisis negotiation, including confidential mandates, public plenary statements, coalition formation, and a convening mediator.

It is designed to preserve information isolation, so that no delegation gains an advantage the corresponding human team would not have had, and to produce a complete and inspectable record of each session, including what each delegation said, which negotiating tactics it used, and how an independent analyst judged the result.

It is designed to express the outcome in variables simple enough for a non-specialist to read, while keeping the reasoning behind each variable visible, and to test how sensitive each outcome is to conditions no negotiator controls, by re-running each scenario under randomly varied external shocks and moods.

Finally, it is designed to document its own method and its own failures, including the specific point at which an orchestration run broke and had to be repeated.

The tool runs each scenario through six stages. The first five use the same reusable toolchain, held in the repository under the folder named sim; only the source documents change from scenario to scenario. Three of the six stages are ordinary computer code with no artificial intelligence involved at all.

Stage one, reading the source documents. Each scenario arrives as a set of documents: one public scenario brief describing the crisis, plus one confidential instruction file for each country. These are converted from page images into plain text. No judgement is exercised at this stage.

Stage two, building a party profile. One agent is assigned to each country. It reads that country's confidential instructions and nothing else, and distils them into a structured profile. This is the single most important step in the whole pipeline, because the profile is the entirety of what that delegation will know about itself for the rest of the negotiation.

Stage three, writing the public brief. A separate agent reads the shared scenario document and writes the neutral briefing that every delegation will see. It states the situation, lists the issues formally on the table, sets out the procedure, and fixes the vocabulary of negotiating tactics that delegations will use to label their own moves.

Stage four, the negotiation. The delegations negotiate in plenary rounds. In each round, every delegation writes exactly one statement. It is given its own profile, the public brief, and the running public transcript. It is instructed to stay in role, to ground its claims in its brief, to protect its secret bottom lines, and to label the tactics it has just used. The flagship Arctic scenario runs four rounds: an opening plenary, a positioning round, a bargaining and coalitions round, and a closing plenary. The other three scenarios run three rounds, dropping the separate positioning stage. Seven delegations across four rounds produce twenty-eight statements in the Arctic; nine delegations across three rounds produce twenty-seven in the South China Sea.

Stage five, the analysis. A single analyst agent, cast as a neutral control group, then reads the entire transcript together with every delegation's confidential profile. It is the only agent in the system that sees both sides of the information barrier. It produces the scoreboard, the per-delegation debriefs, and the convener report.

Stage six, assembly and publication. Plain computer code stitches the outputs together, standardises the tactic labels, and renders the interactive pages. No model is involved. This matters for auditing: the numbers a reader sees on the published pages are arithmetic performed on agent outputs, not a second layer of interpretation.

Everything the simulator produces rests on a small number of variables, each of which has a plain meaning. When an agent reads a country's confidential instructions, it fills in eleven fields. Together they constitute the delegation's entire identity.

Delegation role. One sentence describing who this delegation formally is and how much authority it holds. The Russian profile in the Arctic scenario, for example, records a foreign ministry delegation mandated to defend Arctic sovereignty while deferring major deviations to the Foreign Minister. This field exists because a negotiator's freedom to concede is itself a variable, and a delegation that must telephone home behaves differently from one that can sign.

Fundamental principles and desired end states. The standing commitments the government would assert regardless of this particular crisis give the agent something to argue from when the transcript moves somewhere its specific instructions did not anticipate. End states are recorded as a primary and an alternate. This is the delegation's definition of victory, and its definition of an acceptable substitute. Having two rather than one is deliberate. A negotiator with only a maximum position cannot trade.

Key positions and red lines. The government's stated line on each numbered issue is stored issue by issue, which keeps a delegation consistent across rounds and prevents it from quietly abandoning a position it took an hour earlier. Red lines are the commitments the delegation states it will not cross. In negotiation practice a red line is a declared non-negotiable limit whose value depends entirely on whether other parties believe it. The simulator makes this measurable, because the analyst later records how many of a delegation's red lines were crossed by others.

BATNA. This is a term of art from negotiation theory, introduced by Roger Fisher and William Ury in Getting to Yes in 1981. It stands for Best Alternative to a Negotiated Agreement, and it means the outcome a party falls back on if the talks collapse. It is the benchmark against which every proposal should be judged: a rational party accepts nothing worse than its BATNA. Recording it explicitly is what allows an agent to walk away credibly. Russia's Arctic BATNA, for instance, is to proceed unilaterally with resource extraction and sea-route regulation backed by its partnership with China. A delegation with an attractive fallback has little reason to concede, and the simulator makes that logic explicit rather than leaving it to chance.

Concessions, coalitions, style, and private instructions. What the delegation is authorised to give away determines what can be traded; without it a negotiation of principled statements never becomes a negotiation of packages. Coalition leanings record which other delegations this one expects to work with or against, which allows blocs to form early rather than emerging only by accident. Negotiating style records how the delegation speaks, whether legalistic, conciliatory, or blunt, and tone is part of what is later being judged. Private instructions are the secret material the delegation must never state verbatim, and that field is the reason information isolation matters. If it were shared, the exercise would collapse into a game of open cards.

Every statement in the transcript is stored with six pieces of information: which delegation spoke, which round it was, what kind of statement it was, the full text, the list of tactics the delegation applied, and the list of other delegations it aligned itself with. The last two are what turn a wall of prose into something that can be counted.

Delegations label their own moves from a fixed list of seventeen terms, set in the public brief. The Arctic list runs: anchoring, counter-anchoring, conditional-offer, red-line-signaled, verification-demand, deadline-pressure, coalition-building, issue-linkage, appeal-to-law, appeal-to-precedent, sovereignty-assertion, freedom-of-navigation-frame, side-payment, delay-tactic, principled-bargaining-frame, environmental-frame, and indigenous-inclusion-frame.

Several of these are established concepts rather than inventions of the project. Anchoring is the practice of stating an extreme opening figure in order to drag the eventual settlement toward it, and it draws on the anchoring effect documented by Amos Tversky and Daniel Kahneman in Science in 1974, who found that an arbitrary starting number measurably shifted people's later estimates even when they were paid to be accurate. Issue-linkage means refusing to settle one question except as part of a package with another. A side-payment is a benefit offered outside the disputed issue to buy agreement on it. Principled-bargaining-frame refers to the approach set out in Getting to Yes, which urges parties to argue from interests and objective criteria rather than from fixed positions.

Because agents do not always use the exact word from the list, a short standardising table in the assembly code folds synonyms into the canonical term. Logrolling, package-linkage, and linkage all become issue-linkage. Batna-signaling and red-line-reaffirmation both become red-line-signaled. Face-saving-formula and consensus-appeal both become principled-bargaining-frame. This is a small piece of housekeeping with a real effect on the counts, and it is worth knowing that it happens.

Each tactic label is stored as a detection record carrying a fixed confidence value of 0.8 and a source marked as self-tagging. That number is not a measurement. It is a constant, recording the fact that the label came from the speaker rather than from an independent observer. A reader should treat the tactic counts as a description of what each delegation said it was doing.

The resulting counts are informative in aggregate. In the Arctic session, coalition-building was tagged twenty-five times and red-line-signaled sixteen, against a single appeal-to-precedent. In the South China Sea session, coalition-building appears twenty-four times and conditional-offer fourteen, with delay-tactic appearing only twice. The shape of a negotiation is legible in these numbers before anyone reads a word of the transcript.

The analyst agent then produces four numbers for each delegation. Satisfaction, on a scale from 0 to 1, expresses how well the analyst judges that a delegation achieved the objectives recorded in its own profile. It is scored against that delegation's stated end states, not against any external standard of a good outcome, which is why a delegation can score well in a session that produced no agreement at all. In the Arctic session, Denmark scored 0.75 and Russia 0.38. Agreements won is a simple count of the understandings the delegation secured, and Canada won six in the Arctic session, the highest of the seven. Red lines crossed counts the delegation's declared limits that other parties breached, and Russia's three in the Arctic session, against zero for Denmark, is the clearest single indicator of why the two scored so differently. Self-rating is a separate figure the delegation assigns to its own performance in its debrief. The United States gave itself 3.3 while the analyst assigned it a satisfaction of 0.6. Keeping the two apart is a useful discipline, since the gap between how a delegation rates itself and how a neutral observer rates it is itself diplomatically interesting.

Each debrief also lists the delegation's goals individually. Every goal carries a priority, either critical or high, a status of achieved, partial, or failed, and a short list of evidence quoting the rounds in which the outcome was settled. This is the part of the output that resists being reduced to a number, and it is the part a reader should check before trusting the number.

Alongside the scoreboard, the analyst writes a neutral summary in the voice of a Special Representative of the Secretary-General, the title given to a senior envoy appointed by the United Nations Secretary-General to convene and mediate. The report carries a headline, a summary, a list of key outcomes, and a list of unresolved issues. The Arctic report concluded that the conference produced a scaffolding of cooperation tracks but no binding settlement, and observed that what broke through was the low-politics agenda, where trust was cheapest, while everything touching sovereignty, jurisdiction, and membership stalled.

The published site replays each session as a negotiation table, with a scenario picker, the full transcript, and the session insights attached to each statement. A separate dashboard page presents each delegation's private reasoning, its proposals, the tactic labels attached to its statements, and the scoreboard, for any chosen scenario.

A third page streams a completed session statement by statement, at adjustable speed, to convey the pace of a plenary. The page states explicitly that this is a replay of a finished simulation and that nothing is being generated in the reader's browser.

Each scenario also has a standalone written report combining the overview, the convener report, the scoreboard, and the complete transcript, for readers who want the record rather than the interface.

The Arctic. Seven delegations, four rounds. The issues are the overlapping continental shelf claims of Russia, Canada, and Denmark over the Lomonosov Ridge, which remain before the United Nations Commission on the Limits of the Continental Shelf; the contested legal status of the Northwest Passage and the Northern Sea Route, which Canada and Russia respectively treat as waters under their own control and which the United States regards as international straits; the equal-access provisions of the 1920 Svalbard Treaty and the fisheries zone Norway maintains around the archipelago; militarisation; resources; environment; governance; and the inclusion of Indigenous peoples. The scenario notes that the United States is not a party to the United Nations Convention on the Law of the Sea, which shapes what legal arguments it can and cannot make.

Central Asia and the Fergana Valley. Seven delegations. The issues are the enclave-riddled borders inherited from Soviet administrative divisions, the sharing of Syr Darya water between upstream and downstream states, hydropower, and the competing influence of outside powers.

Cyprus. Six delegations. The issues are the choice between a bizonal bicommunal federation and a two-state settlement, territory, property, security guarantees, and the disputed gas fields of the eastern Mediterranean. This is the only scenario whose source pack predates the others, and it is the scenario the simulator finds hardest to settle.

The South China Sea. Nine delegations, the largest table. The issues are the competing maritime claims, the application of the Law of the Sea Convention, freedom of navigation, and the long-running attempt to agree a code of conduct between China and the states of the Association of Southeast Asian Nations.

A Monte Carlo simulation estimates the range of possible outcomes by running a model many times with randomly drawn inputs and counting how often each result occurs. The technique was devised at Los Alamos in 1946 by Stanislaw Ulam and developed with John von Neumann and Nicholas Metropolis, who supplied the name; Metropolis later recorded that it referred to an uncle of Ulam's who kept borrowing money because he just had to go to Monte Carlo.

A single run tells one story. Whether that story was inevitable or a fluke is a different question, and it is usually the more useful one. Saying that a deal fails in two runs out of eight, and specifically when a hardline shock coincides with a confrontational mood, tells a reader far more about fragility than any single result can.

Each trial draws a fresh value for four things no delegation controls. External shock is drawn from a list written for each scenario. The Arctic list includes a record ice-free summer opening new routes, a Russian naval incident involving a NATO vessel, and a major hydrocarbon discovery in contested waters. The Central Asian list includes a severe drought collapsing Syr Darya flows and a deadly border clash. The South China Sea list includes a collision at Second Thomas Shoal and the declaration of an air defence identification zone. One option on every list is that no major shock occurs, so that the absence of a crisis is itself a sampled condition. Mediator pressure is set to low, medium, or high, representing how hard the convening envoy pushes for a result. Overall mood is set to cooperative, neutral, or confrontational, representing the atmosphere in the room, which experienced negotiators treat as a real variable rather than a soft one. Momentum records which delegation enters with the initiative.

An agent simulates the outcome under that specific draw and reports the type of deal reached, each party's satisfaction, whether each party's primary goal was achieved, partial, or failed, whether its red line held, and which coalitions formed. Deal types are recorded on a five-point ordered scale: comprehensive, framework, partial, stalemate, breakdown. Ordinary code, with no model involved, then counts the deal types, computes each delegation's satisfaction as a mean, a standard deviation, a median, and a lowest and highest value, calculates the percentage of trials in which each red line held, and ranks the coalitions by how often they recurred. A wide spread means the outcome depends on conditions; a narrow one means it is robust.

Eight trials were run for each scenario. Across every trial of every scenario, not one produced a comprehensive settlement. Central Asia was the most stable, returning a framework agreement in six trials of eight. The Arctic returned a framework in five and a partial in three, with no failures at all. Cyprus was by a wide margin the most fragile, returning two outright breakdowns and one stalemate against three frameworks. The Strait of Hormuz session likewise produced two breakdowns. Individual variables are equally revealing: in the Arctic trials, China's red line held in only thirty-eight per cent of runs and the United States red line in eighty-eight per cent, while the five other delegations held theirs in every single trial.

The methodology page is explicit that each trial is a reduced-form outcome simulation conditioned on the random draw, not a complete re-run of the multi-round negotiation. What the exercise samples is the model's own distribution over plausible outcomes. It is not an empirical distribution of real events, and the randomised conditions are illustrative rather than calibrated probabilities. The right reading is how sensitive the model thinks this outcome is to shocks and mood, and nothing stronger.

The variables are not arbitrary. Recording a BATNA, red lines, and a set of tradeable concessions for every party reproduces the standard analytic apparatus of the Harvard Negotiation Project, and it is what allows the simulator to distinguish a delegation that conceded from a delegation that had nothing to gain by holding out.

The distinction between claiming value and creating value, developed in the study of labour negotiations by Richard Walton and Robert McKersie in 1965, appears in the structure of the tactic vocabulary. Anchoring, counter-anchoring, and deadline-pressure are moves that divide a fixed quantity. Issue-linkage, side-payments, and conditional offers are moves that enlarge what is available to divide. The counts therefore say something about what kind of negotiation took place, not merely how much of it there was.

One well-established feature of real diplomacy is deliberately absent. Robert Putnam's account of two-level games, published in International Organization in 1988, describes negotiators bargaining simultaneously at an international table and a domestic one, where a leader whose parliament will reject a deal has less room to agree but more leverage to refuse. The simulator captures a trace of this in the delegation role field, which records how much authority a delegation holds, but it does not model the domestic table. No agent faces a legislature, a coalition partner, or an election. This is a substantial simplification of what constrains real negotiators.

The convening role is modelled at a similarly light touch. The public brief provides that a Special Representative chairs the plenary and may invite proposals and summarise, and the analyst writes in that voice afterwards. But no agent plays the mediator during the talks. Mediator pressure appears only as a randomised condition in the Monte Carlo layer, never as an actor making choices in the room.

The scenario packs are adapted from crisis-negotiation exercise materials associated with the International Strategic Crisis Negotiation Exercise programme run by the Center for Strategic Leadership at the United States Army War College, which takes the exercise to universities each year and has developed regional scenario sets including the Arctic, Cyprus, and the South China Sea.

The repository preserves the source documents themselves, organised by scenario, with one public scenario brief and one confidential instruction file per country. Their file names record the institution and academic year of the exercise from which each pack came, which is an unusually clean audit trail for material of this kind.

The material is notional exercise content written for teaching. It does not represent the actual negotiating position of any government, and the simulator inherits that status. Search-engine indexing of the published site is disabled, both through a site-wide instruction file and through per-page tags, so that AI-generated statements attributed to real states do not circulate as though they were reporting.

The repository documents its own limitations at length, and they are serious enough to state in full.

One mind plays every side. All delegations are driven by the same underlying model. Information is isolated, but the adversaries are not genuinely independent actors, and shared assumptions or style can leak across roles. This is the deepest limitation in the design, because it undercuts the premise that a negotiation is a meeting of different minds.

There is no ground truth. Nothing in the tool has been validated against a real negotiation, an expert assessment, or a historical outcome. The scoreboard numbers are one system's subjective judgement of another system's writing. Satisfaction should be read as how the analyst rated the argument on the page, not as an assessment of any national interest.

Agents can fabricate. A language model can invent a plausible-sounding treaty, figure, or precedent that appears nowhere in its brief, and can drift from a government's real position in ways that read smoothly.

Training data shapes the outcome. Which arguments sound strong, which framings feel reasonable, and how sympathetically each actor is portrayed all reflect biases in the material the model learned from. A simulator in which one delegation consistently argues more persuasively than another may be reporting a fact about the model rather than about the dispute.

The exercise is compressed. A multi-day negotiation becomes three or four short plenary rounds. Caucuses, bilateral side-channels, corridor conversations, and escalation dynamics are simplified or absent. Much of what determines the outcome of real talks happens outside the plenary, and none of that is here.

The samples are small. Eight trials per scenario is enough to show that Cyprus is fragile and Central Asia is not. It is not enough to support a percentage figure quoted on its own.

One accounting discrepancy. The methodology page describes the Monte Carlo run as thirty-two trials, counting the four multi-party scenarios. The aggregated data file also holds eight trials for the earlier Strait of Hormuz session, giving forty in total. Neither figure is wrong, but they count different things.

The project records a failure honestly. The first orchestration run struck a transient network outage that stopped roughly half its agents mid-negotiation, and a targeted re-run completed the missing rounds. The repository states that no gaps were papered over and that every published statement is a genuine agent output, with the intermediate artefacts committed for inspection. Because language models are not deterministic, the exact wording of any repeat run will differ even from identical source documents.

It is not decision support. The repository states directly that the tool must not be used to inform real policy, negotiation, or intelligence judgements.

The prototype is built for teaching and for structured exploration. Its most defensible use is as a companion to a human exercise rather than a replacement for one. A class that has run the Arctic scenario itself can compare its own transcript against the simulated one and argue about where the agents were wrong, which is a more demanding exercise than either activity alone.

It is also of interest to researchers examining what happens when language models are asked to hold a position under pressure. The run artefacts are complete enough to support that kind of scrutiny, which is precisely what makes the honest limitations section valuable rather than decorative.

It is of no use to anyone seeking to know what any government will actually do. The tool says so itself, and this report repeats it because the output is fluent enough to be mistaken for analysis.

The Diplomatic Simulator is a modest and carefully documented prototype. Its contribution is less the simulated negotiations themselves than the discipline imposed on them: confidential mandates that stay confidential, statements that are counted and labelled, scores that are defined against each delegation's own stated goals rather than an invented standard of success, and a sensitivity layer that reports how much of the outcome was luck.

The most valuable thing in the repository may be its methodology page, which sets out what the agents do, what they do not do, which model produced each layer, where the run broke, and why none of the numbers carry ground truth. Tools of this kind will become common, and their outputs will be fluent enough that readers extend them more credibility than they have earned. A prototype that publishes its own failure modes alongside its results, and that turns off search indexing so its invented statements are not mistaken for reporting, sets a standard worth borrowing.