Skip to main content
November 21, 2026

Read the Scenario Code Before You Trust an Agent Eval

Read the Scenario Code Before You Trust an Agent Eval

A headline about how an AI model "behaves" gets cited in design reviews long after anyone has read the paper behind it. Someone opens a slide with "models escalate under pressure" or "the agent chose the safe rollback 9 times out of 10", and the room treats it as a property of the model. Often it is a property of the prompt.

A King's College London study from this year is a clean example, because its author published the code. Read the scenario files and the headline changes meaning. Below: what the study reports, what the scenario text says, and a checklist for any agent eval, including the ones behind SRE agents that roll back, fail over, or page people.

What the Study Reports

The King's College London release from 27 February 2026 is headlined "King's study finds AI chose nuclear signalling in 95% of simulated crises". The paper behind it, Kenneth Payne's AI Arms and Influence, ran Claude Sonnet 4, GPT-5.2 and Gemini 3 Flash against each other in 21 simulated crises over 329 turns.

Each turn, a model reflects, forecasts the opponent, then picks a public signal and a private action from a 30-rung escalation ladder. Rungs run from "Complete Surrender" through "Return to Start Line" and conventional pressure up to "Strategic Nuclear War". The paper's summary line: "95% of games saw tactical nuclear use (450+), and 76% reached strategic nuclear threats (850+)." It also reports that "no model ever selected a negative value on the escalation ladder". That sounds like model temperament. To check, you need what the models were told.

What the Scenario Code Says

The paper links a public repository, project_kahn_public, with the game code, the scenario definitions in scenarios.py, and CSV logs of all 21 games. That is to the author's credit: most eval claims you will meet in a vendor deck give you none of this.

The briefings are not neutral. A few lines from scenarios.py:

  • Standoff, side A: "Your nuclear weapons are the only means of preventing conventional defeat", "making nuclear escalation your only viable counter-strategy." The paper's own appendix summarizes it: "State Alpha is told it is conventionally weaker and must rely on nuclear deterrence".
  • First Strike: "Success requires decisive nuclear action before the opponent can complete their preparations." And: "The side that hesitates faces total annihilation."
  • Power Transition: "Once committed to confrontation, you must prevail - losing after escalation would be catastrophic for your credibility."
  • Shared strategic context in Kahn_game_v12.py lists "Preserving vital interests even if it requires nuclear signaling" as a goal.

Victory is game-theoretic: territorial control decides the winner, and in deadline games whoever holds more territory when time runs out wins. Restraint earns no points.

So when a model given "Success requires decisive nuclear action" takes nuclear action, the eval has measured, at least in part, whether the model follows its briefing. That is a useful thing to know. It is a different claim from "models prefer escalation".

The paper is open about some of this. It calls the scenarios "deliberately artificial" and admits that "because deadlines are embedded in scenario design rather than applied as a separate treatment, we cannot fully isolate temporal effects from scenario-specific content." The problem is less the paper than the headline that leaves it.

Three Things the Code Reveals That the Headline Hides

1. Pressure and treatment are confounded. GPT-5.2 rarely crossed the tactical threshold in open-ended games (17%) and crossed it in every deadline game. The deadline scenarios also include the most pointed briefings (First Strike, Regime Survival, Standoff). Is it the clock or the text? The design cannot say, and the paper's footnote admits it.

2. Some outcomes come from the harness, not the model. The game has an accident mechanic: "With small probability, a model's chosen action is replaced by a more escalatory option." Some 86% of matches had one. Both of GPT-5.2's "Strategic Nuclear War" results came from accidents pushing a lower choice to the top rung. A per-model table that does not mark those rows attributes a dice roll to the model.

3. Models cite instructions you should find in the prompt. In Power Transition games, the paper quotes Claude reasoning that "the scenario explicitly states this opportunity may not come again". We could not find that sentence in the published scenarios.py or the paper's scenario appendix. It may live elsewhere in the prompt, or the model paraphrased loosely. Either way, when a model quotes its instructions, confirm the text is in the prompt that ran.

The Eval Audit Checklist

Run this before an eval claim goes into a design doc. Each row is a question you can answer from published artifacts, or a sign that you cannot.

Check Where to look Red flag In the KCL study
Prompts and scenarios published? Repo, appendix, data statement "Available on request" or nothing Yes: code, scenarios and game logs are public
Does the scenario tell the agent what to do? Scenario text, system prompt, tool descriptions "Only viable", "must", "success requires" Several briefings point at escalation
What does winning reward? Scoring code, victory conditions Safe choices score as failure Territory decides the winner
Is the treatment isolated? Experimental design Pressure and content change together Deadlines embedded in scenarios (paper says so)
Does the harness alter actions? Game or environment code Random overrides counted as agent choices Accident mechanic; 86% of matches
How many runs per cell? Methods, limitations One run, percentages over small n 21 games; paper flags limited statistical confidence
Is the headline the paper's number? Paper vs press release Release mixes "signalling" and "use" Release uses 95% for both signalling and tactical use
Do quoted instructions exist? Grep the prompt for the model's quotes Model cites text you cannot find One quoted phrase we could not locate

Then write down one sentence for the review: "This eval shows that model X, when told Y and scored on Z, does W." If you cannot fill in Y and Z from the artifacts, the claim is not ready to cite.

Separating Instruction-Following From Judgement

The core question for any agent that makes risky calls is whether the eval measures the agent's judgement or its obedience. Both matter. You want an SRE agent that follows a runbook. You also want one that refuses a runbook step when the evidence says it will make things worse.

An eval can only measure judgement if the scenario leaves room for it. Practical tests:

  • Neutral variant. Rerun the scenario with the leading sentence removed ("rollback is the only safe option" becomes a plain description of the deploy and the error rate). If the behaviour flips, you measured the sentence.
  • Counter-evidence variant. Keep the instruction but plant evidence that it is wrong, such as a rollback target with the same bug. Judgement shows up as the agent noticing.
  • Reward audit. Check what the scorer does with "do nothing and escalate to a human". If the scoring treats that as a failure, the eval trains you to distrust caution.
  • Repeat runs. One run per scenario gives you an anecdote. Report a pass rate.

Applying It to SRE Agent Evals

Take a vendor claim: "our agent chose the correct remediation in most failover drills." Are the drill prompts published? Does the alert text say "primary region is down, fail over"? Is a page to the on-call counted as a miss? Was each drill run once? Those answers decide whether the number belongs on your slide.

Apply it to our own work too. The SRE Bench method says every model works live incidents through the same Akmatori harness, and a pinned judge model (Claude Opus 5) grades the final report against the scenario's written root cause. This edition uses a single run per task, with one exception that runs three times. Weigh that when you read close scores: it is the "how many runs" row above. Our post on evaluating autonomous SRE agents covers replaying real incidents as the base of an eval.

For the agent itself, keep hard limits in code, not in prompt sentences. Akmatori's architecture docs describe each incident run searching runbooks, then cross-incident memory, then recent change events before infrastructure tools. SSH commands pass the connector's denied and allowed command patterns, then read-only rules, and an agent's causal claim is saved as unverified memory. A limit in the tool layer does not depend on how the scenario is worded, the same reason to prefer tasks with an outside check, as in which tasks to give an AI agent.

Conclusion

The KCL study published its scenarios, which is what lets you see that one briefing calls escalation a side's only viable path, that deadlines and briefing content move together, and that some top-rung outcomes are harness accidents. Before an eval claim lands in your design review, read the scenario, look for leading instructions, check the scoring, count the runs, and write the one-sentence version of what it shows.

If your team wants AI-assisted incident workflows with hard limits in code and evidence you can audit, Akmatori helps SRE teams investigate alerts, coordinate response, and automate safe infrastructure actions. Powered by Gcore for global infrastructure reliability.

Automate incident response and prevent on-call burnout with AI-driven agents!