Robustness

For Both Executive and Technical Readers

A posterior is a point. The question a regulated decision turns on is how fragile that point is: to a wrong expert estimate inside the model, and to a wrong or missing observation feeding it.

The model answered the question. It didn’t say how much you should trust it.

Most probabilistic systems answer the question they were asked and stop. They do not tell you whether their answer would change if an input were slightly wrong. In a regulated context, that omission is the vulnerability: a conclusion built on a single shaky input that nobody tested.

A conclusion can be wrong two ways. The parameters can be off: an expert’s probability estimate inside the model is slightly wrong. The evidence can be off: an observation feeding the model is missing, mistaken, or about to be revised. Robustness analysis tests both.

Two analyses, sensitivity to assumptions, and necessity of evidence.

Robustness is a Cognitive Primitives, Epistemics component. It runs two complementary analyses on any inference result: retracted analysis, asking what the model would conclude if a piece of evidence turned out to be wrong or absent, (what if this observation were wrong or absent?) and sensitivity analysis (what if this parameter estimate were off?). Together they produce a RobustnessReport: a typed artifact showing which conclusions are solid and which rest on a single shaky input.

Robustness Report two analyses, one artifact
AnalysisWhat it answers
Retracted evidenceFor each piece of evidence, what the posterior would be without it, and how far it moves, so you can see which observations the conclusion actually leans on
SensitivityFor each parameter, the range the posterior spans if that number were off, and whether that range is wide enough to flip the conclusion
Verdict: robust, fragile, or conditionally robust. One word, always present, never omitted.

The verdict field is machine-readable: a downstream governance artifact can require "robust" before a decision proceeds without human review.

Perturb the CPTs. Re-run. Measure what changed.

Two analyses, run on the current inference result:

Retracted analysis. Remove each piece of evidence in turn and re-run inference. A positive blood culture pushed sepsis posterior to 0.93, remove it, and the posterior falls to 0.66. The conclusion leans heavily on that one result. Remove the temperature reading, and almost nothing moves. Now the clinician knows exactly which inputs the diagnosis depends on.

Sensitivity analysis. Perturb each expert parameter estimate across a plausible range and observe the posterior range. If P(Shock | Sepsis) = 0.73 could be anywhere from 0.60 to 0.85, and the conclusion barely moves, the estimate can be rough. If it flips the decision, that is the parameter most deserving an expert’s care before production.

Verdict. The RobustnessReport combines both into a single verdict, auditable by a regulator: which inputs the conclusion depends on, and how much it would take to change it.

The model checking its own work.

Robustness is the brains checking its own work, testing which of its conclusions would survive challenge and which would not.

There is no .bayes file for Robustness, Robustness operates on any domain model. The artifact it produces is a RobustnessReport attached to the AuditRecord. In regulated contexts, the report is a required output alongside the posterior.

Robustness Report sepsis_model · worked example
AnalysisResult
Retracted evidenceRemoving BloodCulture: posterior drops to 0.66, a delta of −0.27, the conclusion leans heavily on this observation
SensitivityP(Shock | Sepsis): posterior ranges 0.81–0.85 across plausible values, not critical, this parameter can be rough

How much should I trust this answer?

Robustness answers the question that comes after the posterior: how much should I trust it?

P(Sepsis) is 0.93. Is that number solid?
AnyChatRobustness · retracted analysis
Running retracted analysis. Remove the blood culture: posterior falls to 0.66, the conclusion depends heavily on that one result. Remove the temperature: almost no movement (0.92). The 0.93 is fragile to the blood culture. If that result were revised or in doubt, the call changes.
What about the model parameters themselves?
AnyChatRobustness · sensitivity analysis
Running sensitivity analysis. The parameter most worth getting right is P(Sepsis | Lactate>4, Age>65), perturbing it across the expert’s confidence interval moves the posterior from 0.88 to 0.97. That is the estimate most deserving scrutiny before production. All other parameters are below the critical threshold.
How do you know which inputs to test?
AnyChatProvenance
I don’t, the model does. Retracted analysis removes each EvidenceBinding in turn and re-runs exact inference. Sensitivity analysis perturbs each parameter across a specified range. Both are deterministic operations. The RobustnessReport is attached to the AuditRecord.
Couldn't our risk team just run Monte Carlo on the model inputs?
AnyChatRung 2 · sensitivity
Monte Carlo samples from assumed input distributions and propagates forward. It does not tell you which inputs the output is sensitive to, or whether the model’s parameters themselves are the fragile point. Robustness analysis asks a different question: how much does the conclusion change if this CPT is wrong by 20%? Which node, if its prior shifted, would flip the decision? That is a structured sensitivity sweep over the model’s own uncertainty, not over data inputs. The answer determines where elicitation effort should go next, and whether the current model is stable enough to act on.

The audit trail is the RobustnessReport, not the analyst’s judgment about which inputs matter.

Downstream of Execute, upstream of the governance handoff.

Robustness sits downstream of Execute (where the posterior is computed) and upstream of the governance handoff (12 BPMN Integration), which may require a "robust" verdict before the pipeline proceeds without human review.

Upstream: the inference result and AuditRecord from Execute (01 Pipeline stage 6). Downstream: the RobustnessReport is appended to the AuditRecord and consumed by 12 BPMN Integration. It also informs 07 Validation (calibration checks) and feeds back to 04 Cognitive Primitives (which parameters most need re-elicitation).