Sensitivity and switching values

Sensitivity analysis asks whether the recommendation is robust to plausible changes in inputs. It is a conditional stress test, not a probability distribution and not evidence that every value inside a range is equally likely.

What version 1 automates

The built-in workbench automates only one-way chance-probability sensitivity. Select one outgoing branch of a chance node and define a defensible minimum, maximum, and bounded grid. At every sampled probability, version 1 redistributes the remaining probability proportionally across that node's sibling branches so the total remains 1, then rolls back the complete tree and records the affected decision recommendation and expected value. These transient cases do not change the saved model.

The workbench reports an approximate switch interval when the preferred branch differs between two adjacent sampled points. The interval brackets where a change occurred; it is not an exact analytic tie or break-even value. A coarse grid can locate a switch only loosely and can miss behavior between samples, so refine the range and grid when a narrower, evidence-supported bracket matters.

Choose the range from evidence rather than an arbitrary percentage. Record the source, confidence, range, as-of date, and reviewer in the branch's probability basis or linked evidence notes; version 1 has no separate fields for those probability metadata.

What version 1 does not automate

Version 1 does not run sensitivity on consequences or terminal values, branch increments, information costs, or the unacceptable-outcome threshold. It also does not solve exact break-even values, create tornado charts, or run joint scenarios, multi-way sensitivity, or Monte Carlo simulation. These are external follow-up analyses, not built-in results. A report or assistant response must not imply that the workbench executed them.

External follow-up guidance

Outside the built-in tool, consequence, cost, and threshold cases can be useful when their ranges are sourced and the entire policy is recalculated for every case. A tornado chart may summarize comparable one-way analyses, but its ranking depends on both the chosen ranges and model response; it is not causal importance. Related inputs may require coherent joint scenarios or simulation that preserves correlations and conditional structure. Never vary sibling chance probabilities independently and then normalize away an incoherent event model.

Robustness statement

For the built-in analysis, summarize which policy remains preferred across the sampled probability grid, any approximate switch interval, how expected value changes, and where evidence is weakest. If plausible sampled values reverse the decision, describe the alternatives as closely ranked and consider additional information. If they do not, say only that the policy is robust within this tested model, probability range, and grid—not across untested inputs and not that it will succeed.