How the instrument works

The engine in one paragraph

Every simulation starts from a world state: the entities in a domain (patients, orders, employees, portfolios), their relationships, and the rules governing how they interact. A decision or disruption is injected as an event. The engine — a discrete-event simulator — then advances the world forward, event by event, sampling every uncertain input from a probability distribution rather than a single guess. It does this not once but thousands of times in parallel, each run drawing different samples — a Monte Carlo simulation, in the standard sense of the term. The output is never one prediction. It is the full distribution of what could happen: the median, the percentiles, the tails, and the confidence interval on every metric.

One engine, many domains

The engine itself never changes between domains. What changes is the domain schema — a declarative file describing the entities, events, distributions, and constraints of one world. A hospital and a supply chain share the same mathematical skeleton: a central entity flowing through constrained resources, with consequential events cascading through a relationship graph. Fifteen domains currently run live on the same core, from ICU capacity to portfolio drawdowns to grid stability. The domain catalogue lists them with the questions each can answer.

What the confidence intervals mean

When a finding says "lead times breach SLA in 62% of runs (CI 57–66%)," it means: given the stated assumptions, 62% of the simulated futures crossed that threshold, and the interval reflects sampling uncertainty across the batch. The number is conditional on the assumptions — always visible, always editable. If you disagree with an assumption, you are not asked to argue; you are invited to change it and re-run. The disagreement becomes a second run, and the two distributions can be compared side by side.

Calibration: how the model earns the numbers

Distributions are not invented. Each schema declares distribution families for its uncertain inputs — arrival rates, lead times, attrition hazards, return processes — and these are fitted to historical data. Where real-world outcomes are observable, the model is tuned against them using approximate Bayesian computation: parameter sets that reproduce observed reality are kept, those that don't are discarded, and the surviving posterior drives future runs. Calibration is continuous — when reality disagrees with the model, the model changes, and the change is noted in the run lineage.

Whether this works is not a claim you need to take on faith. The track record grades every public prediction against what actually happened, and publishes the calibration score: when the engine says 80%, does the thing happen about 80% of the time?

Reproducibility

Every run is seeded. The same run with the same seed produces the same result, byte for byte — which is what makes the replay link possible. A finding shared with you is not a screenshot of an analysis; it is the analysis, re-executable, with its complete assumption set attached. Runs are never edited after publication. Corrections are new runs, linked to the original.

The mark: what it certifies, and what it doesn't

A verified finding carries the observatory's mark. The mark certifies the math: that the computation ran on the calibrated engine, at the stated scale, from the assumptions shown, and that the result is reproducible from the attached seed. It never certifies the assumptions. Assumptions are judgment — they belong to whoever ran the analysis, and they are always displayed in full precisely so you can judge them. This division is deliberate: the engine guarantees honest arithmetic; the analyst owns the reasoning; and you can see both.

What this instrument cannot do

It cannot predict events its assumptions contain no signal for — a demand model fitted to history will not foresee a structural break history hasn't seen. It cannot rescue a bad question; a precisely simulated answer to the wrong question is still wrong. It does not model strategic adversaries who react to your decision (game-theoretic response is outside the current engine). And a probability is not a promise: an 80% claim is expected to fail one time in five, and a well-calibrated record will show exactly that. Anyone selling you certainty is selling you something other than a simulation.

Who runs this

Decision Observatory is operated as an automated instrument. Runs, reports, and support are delivered by the system itself, and the humans behind it spend their time where it compounds: making the engine more accurate, the domains deeper, and the record more complete. We think an instrument should be judged by its output and its grades, not its founders — which is why the track record is the About page.