Decidability map — page 3 of 9
Each outcome claim the client cares about is scored against one question: can this be proven from your existing data, on your own scoreboard — and at what cost? Verdicts below are the three we issue. NOT PROVABLE is a finding, not a failure — it tells you what to instrument before spending on transformation.
C-01
Auto-clearing supplier invoices reduces processing cost per invoice
ScoreboardAP cost centers in the general ledger; per-invoice cost derivable from accounts
Baseline14 months of booked history, stable process owner
BasisPre/post comparison on the client's own cost-per-invoice line; seasonality controlled by
C-02
Collections agent raises recovery rate on the 30–60 day bucket
ScoreboardBooked recoveries, bucket-level, already reported monthly to
Baseline24 months, bucket migration rates stable
BasisRecovery-rate delta on the regulator-visible number; portfolio mix shift is the named threat and is measurable
C-03
Agentic KYC cuts onboarding time without raising file-error rate
ScoreboardOnboarding-file timestamps + error findings in reviews
GapError-rate ground truth requires an adjudicated sample — it does not exist in telemetry today
Cost to proveSampling protocol on ~ files per quarter, two adjudicators, defined error taxonomy
C-04
Assistant improves customer satisfaction
ScoreboardPost-interaction survey sentiment
FindingOpinion baseline — declined under the scoreboard rule. A verdict on this claim would be an opinion wearing a number
C-05
Fraud-screening agent lowers losses attributable to the agent
ScoreboardBooked fraud losses (objective, auditable)
FindingAttribution confounded today: rule-engine changes shipped in the same windows as
PathBecomes provable after change-windows are separated — instrumentation spec included in section 7
ProvableYour data supports a verdict today.
Provable at costA verdict is possible after named, priced instrumentation.
Not provableWe will not certify it — and we tell you what would change that.