firehose> #llmops

Wargaming (Adversarial Contingency Planning)

A wargame is a plan written for the world not going as planned. A normal plan — even from a very capable model — assumes linearity: it breaks the task into phases, sequences them logically, and reads as a high-probability blue-sky path. A wargame instead has the model fight the mission on paper move by move, and at every move enumerate the failure it could hit, the signal that reveals the failure, and the counter-move that resolves it. The source's spine is three elements — action → reaction → counteraction: the model makes a move, "reality humbles it" by throwing an error, and it must take a counter-action to recover. The source calls this loop, pre-simulated on paper, "the modern-day agentic loop" written out ahead of time (see Agent Loop).

The move that makes it more than a checklist is the explicit instruction "you are not executing this mission, you are purely war-gaming it" — a capable model distinguishes a plan from a wargame and, told which it is producing, front-loads the contingency reasoning it would otherwise discover only at run time. The concrete artifact each move carries: expected observation (what you see if it worked) / most-likely failure + the cause it signals + the counter-move / a fork trigger ("if you observe X, take route B") / RECON NEEDED markers for assumptions due-diligence could not settle (each with the exact check that would settle it) / and terminal abort conditions — the error classes at which the executor should stop rather than push on (e.g. no access to a required system). The operator sets the simulation depth — how many orders of consequence (second-, third-, fourth-order) to game out before stopping — because unbounded simulation never terminates.

Why do this with a premium model you are about to lose access to: the wargame is a portable artifact. Once the hard contingency reasoning is captured in markdown, a cheaper or open-weight executor (the source names Opus 4.8, GPT-5.5, GLM, Sonnet 5) can run the brief "end to end without asking a single question" and far more confidently, because the realities it will meet have already been simulated. This is the "method you feed the model" made durable — see Execution Commoditization and the architect/executor split of Model-Tier Routing. It is distinct from live plan-then-dispatch routing: the expensive model's reasoning is distilled into a transferable blueprint, not consulted in the loop.

Claims


Linked from