Platt calibration of the duplicate Likelihood ("the candidate is the same real-world entity as
the new record"), per decision model, each tied to the exact model it was fitted on (see
FindDecisionCalibration in @memberjunction/ai-core-plus): the MJ decision model that answered
(ModelName) and the model the driver reports behind it (ResolvedModel). A model's raw
probabilities are not calibrated: Jev ranks candidates almost perfectly (AUC 0.99) but its raw
probabilities run high for similar records that are not duplicates.
Any other model, including Jev at another version or LLM Decision answered by another chat
model, gives no calibrated probability, since its raw one can't be banded. The entry check then
flags nothing, as it does for a failed decision; batch Decision mode bands its candidates
Uncertain for review; and DecisionThenPrompt sends them all to the prompt. The provider logs
the missing calibration once per model. An app that binds another decision model supplies its
calibration by overriding DecisionReasoningProvider.CalibrationFor.
Fitted on the duplicate-check measurement (plan Task 3.8, 2026-09-29), on every candidate: 160
labelled new records for MJ: Actions (80 rewrites of an existing action, 80 similar actions that
don't exist), five vector candidates each.
Jev at its pinned APIName, typesafe/jev-1.13-20260917, which OpenRouter reports back as
the resolved model.
LLM Decision when its chat model is GPT-OSS-120B, its LLM Decision prompt's first choice.
The stored runs don't record the resolved model; this is the prompt's binding on that day.
Refit whenever a model's pinned version, its delegated chat model or the question changes.
Platt calibration of the duplicate Likelihood ("the candidate is the same real-world entity as the new record"), per decision model, each tied to the exact model it was fitted on (see
FindDecisionCalibrationin@memberjunction/ai-core-plus): the MJ decision model that answered (ModelName) and the model the driver reports behind it (ResolvedModel). A model's raw probabilities are not calibrated: Jev ranks candidates almost perfectly (AUC 0.99) but its raw probabilities run high for similar records that are not duplicates.Any other model, including Jev at another version or
LLM Decisionanswered by another chat model, gives no calibrated probability, since its raw one can't be banded. The entry check then flags nothing, as it does for a failed decision; batchDecisionmode bands its candidatesUncertainfor review; andDecisionThenPromptsends them all to the prompt. The provider logs the missing calibration once per model. An app that binds another decision model supplies its calibration by overriding DecisionReasoningProvider.CalibrationFor.Fitted on the duplicate-check measurement (plan Task 3.8, 2026-09-29), on every candidate: 160 labelled new records for MJ: Actions (80 rewrites of an existing action, 80 similar actions that don't exist), five vector candidates each.
APIName,typesafe/jev-1.13-20260917, which OpenRouter reports back as the resolved model.LLM Decisionprompt's first choice. The stored runs don't record the resolved model; this is the prompt's binding on that day.Refit whenever a model's pinned version, its delegated chat model or the question changes.