Teaching a simulation to be boring, on purpose.

Language models love a story. Left alone, a simulated negotiation sprouts a whistleblower; a simulated product launch sprouts a scandal; a simulated year sprouts six. That instinct makes great fiction and bad decision support. Here's what we've been building into Timepoint Pro to fix it — and why the most important feature is the one that refuses to hand you the result.

The problem: engines optimized for drama

A simulation engine is, underneath, a storyteller — and storytellers get rewarded for surprise. But when a client asks Pro to play out a real decision, surprise is exactly the wrong objective. The useful output is the typical path, the base-rate path, the path where people act on their actual incentives and the calendar behaves. Drama should appear only when the situation genuinely produces it.

So the work of the last stretch has been what we call grounding: teaching Pro to be faithfully boring, and building the machinery to prove it run by run.

What a grounded run must survive

Every grounded simulation is now audited against six families of checks:

1. Base-rate realism. Events occur at plausible frequencies — no invented crises for the sake of a better scene.
2. Typicality. The run stays inside its reference class. A contract dispute behaves like contract disputes actually behave, not like a thriller about one.
3. Allegiance. Simulated actors act on their real documented incentives — nobody switches sides because the plot wanted it.
4. Time discipline. Ask for twelve months and the simulation paces twelve months — not the multi-year wander early versions preferred. (This was a real bug: a hardcoded pacing window that ignored the instructed horizon. It's fixed, and we found one residual dating quirk we're still closing.)
5. Disposition. Outcomes stay consistent with how the situation was actually disposed to unfold, not with what makes the strongest ending.
6. Prohibitions. Things the scenario rules out stay out. Period.

Feeding those checks is a new grounding source: before a run, the engine builds a reference-class brief — what situations like this one typically look like, derived from the facts of the entities involved — and the branch scorer now rewards staying true to it. On grounded runs, the scoring boosts that reward narrative fireworks are switched off entirely.

The gate that says no

Checks alone are advisory. The part that changes what clients receive is the enforcement gate: a run that commits a hard violation is blocked — it does not become a deliverable, full stop. In our latest acceptance testing, across every run in two full rounds, the gate caught what it was supposed to catch and zero violations shipped.

The property we care most about is what happens when the checking machinery itself has a bad day. Midway through testing, an upstream model provider had an outage — the exact moment a lazier system would shrug and pass everything through. Pro's audit fails closed: when it can't evaluate a run, the run is marked unevaluated, never silently passed. The system would rather tell you it doesn't know than pretend it checked.

What this means if you're a client

What reaches you has survived the gate. A Pro deliverable now carries its own audit trail: which checks ran, what they found, and whether anything was blocked and why. When a run can't be evaluated, the deliverable says so in plain sight — you will never get a vacuous "all clear."

The horizon you ask for is the horizon you get. Decision work has a clock: a board meeting, a filing window, a renewal date. Time discipline means the simulation spends its effort inside your window instead of wandering years past it.

Boring findings you can act on. The point of neutralizing drama isn't blandness — it's that when a grounded run does surface a crisis branch, it earned its place. That's what makes the blind-spot findings in our client work worth acting on: they survive checks designed to kill invented ones.

Most of the industry is teaching simulations to be more impressive. We think the harder and more valuable engineering is teaching one to be trustworthy — boring where reality is boring, dramatic only where it's earned, and honest about its own blind spots the rest of the time.